fix(compilers/openapi): keep path-item docs and extra operations

[OmarAlJarrah]

Summary

The Path Item Object loses more of itself than any other object in the OpenAPI compiler.
lowerPathItem reads the fixed method fields, parameters, servers and $ref, and nothing
else, so three of the object's remaining fields reached the IR in no form at all — no field, no
Unmodeled entry, and no diagnostic on either channel:

• summary / description.fillOperationDocs reads the Operation Object's own pair, so a
  path item's documentation was dropped outright.
• additionalOperations (3.2). A map from an HTTP method with no fixed field of its own to a
  full Operation Object. Every operation declared there vanished whole — operationId,
  parameters, request body, responses — and with it every type reachable only through it, so the
  type registry came out short as well.
• query (3.2). The fixed field 3.2 added for the QUERY method, missing from the accessor
  table the walk iterates. It is the same whole-operation loss as additionalOperations at the
  same site, so it is fixed here rather than left to reopen this function for one table entry.

The loss repeated at all three routes that reach a path item — paths, webhooks, and a callback
expression. Neither issue says so, and a fix scoped to lowerPathItem alone would have missed two
of the three.

Where the documentation lands

summary and description are now kept verbatim on each operation the path item holds, under
openapi:pathItemSummary / openapi:pathItemDescription with no_ir_home and one info
diagnostic, rather than merged into Docs.

There is no ir.PathItem — a path item is distributed across the operations it holds — so the two
candidate destinations were a merge into Operation.Docs or preservation beside it. ir.Docs
holds one summary and one description and they are the operation's own, while a path item's pair
documents the path. Merging would need a precedence rule against the operation's own, would attach
to an operation documentation its author never wrote (an inference, which belongs in injectable
policy rather than in a lowering), and would leave an emitter unable to tell the two subjects
apart afterwards. Preserving takes no position on precedence and loses nothing. The cost — the
pair is duplicated onto every operation under the path — is the cost path-item servers already
pays at the same site. The reasoning is recorded on applyPathItemDocs and in docs/ir-design.md,
and no_ir_home rather than a boundary reason because the IR could grow a home for it.

How the operations lower

additionalOperations entries become ordinary ir.Operations mounted at their own pointer
(…/additionalOperations/<method>), which is where the source writes them, so none takes an ID a
fixed method field could also claim. The map key becomes HTTPBinding.Methodverbatim: that
field is documented as the method as sent on the wire and OpenAPI reads a method name
case-sensitively, so the compiler neither upper-cases nor neutralizes it. A fixed field's name is
a field name rather than a method, so that one is still upper-cased into its wire spelling.

The three walks now share one pathOperations(pi) reader, and a path item's servers and
documentation are kept through one applyPathItemResidue call per route, so the next construct in
either class cannot reach two routes and miss the third.

Not in scope

A path item's x-* extensions are still dropped. That is a different mechanism — every object
that admits an extension and has no reader for it — and is tracked in #275, which covers the
dozen-odd objects sharing it rather than this one.

Two cases keep the documentation in no form, both shared with the servers half beside it
rather than introduced here. Each is called out on applyPathItemDocs:

• A path item that mounts no operation. The pair is kept on each operation the item holds,
  so an item holding none keeps nothing. openapi: a path item's summary and description reach the IR in no form #383 records that gap for the whole class, and the
  unmounted-path-item work is where it lands.
• A pair supplied through a YAML merge key or an alias.RawChildNode scans a mapping's
  pairs directly, so the model reads the pair and the lookup does not: GetSummary is
  non-empty, the raw node is nil, and PreserveNode writes nothing and returns no diagnostic.
  Filed as openapi: a preserved field supplied through a merge key or alias is dropped with no trace #384 against the lookup rather than this caller — 31 non-test sites share it, and
  applyPathServers has had it since openapi: several source details silently dropped, violating lossless-by-default #39.

One key that names no method

additionalOperations keys reach HTTPBinding.Method verbatim, which is what an empty key
needs reporting for: it binds a method no request can be sent with, and speakeasy accepts it
(it rejects only a key naming a standard method, which has a fixed field of its own). The
entry still lowers in full — dropping it would trade a reported defect for a silent one — and
openapi/invalid-method-key names the entry that declared it. The check sits in
lowerOperation, where all three routes funnel through.

Test plan

• New tests in compilers/openapi/internal/operation: the documentation pair and an
  additionalOperations entry are each asserted at all three routes with distinct text and
  distinct operationIds apiece, so a fix that kept the enclosing path item's — or the same one
  three times — fails rather than passing on the shape alone. The query field, verbatim method
  casing, the interning of a type reachable only through a previously dropped operation, and a
  callback expression that declares no fixed method at all are asserted alongside.
• Two conformance fixtures with goldens: path-item-docs (3.1) and path-item-operations (3.2),
  each covering the path, webhook and callback mounts in one document, so the corpus sweeps —
  irverify, the JSON round-trip, the determinism and two-order oracles — run over both.
• Verified the corpus bites: deleting the path item's summary line from one fixture and the
  mixed-case additionalOperations entry from the other reddens TestConformance for both.
• Verified the tests bite: with the change to operations.go reverted and the tests kept, the
  three new behavioural tests go red (postP keeps its path item's summary, operation "queryP" not found, operation "purgeCallback" not found) and both conformance rows fail.
• Full gate green: gofmt, go vet, golangci-lint, go build, and the coverage gate at
  exactly 100%. No existing golden moved.

Closes#292
Closes#293

Fixes the mounted half of #383, which spans a path item whether or not it mounts an
operation. The unmounted half is untouched — there is no operation to carry the entry —
so #383 stays open for the unmounted-path-item work rather than being closed here.