refactor: tighten the public API and name the placeholder-style test

[iamralch]

Follow-up to a read-through for idiomaticity. clippy::all + pedantic were
already clean, and nursery + cargo turned up only redundant_pub_crate and
dependency-version noise from sqlx's tree — so these came from reading, not
from lints.

Dimension is no longer #[non_exhaustive]

An AIP List request has exactly these three query dimensions, and the crate's
scope excludes taking on more. Marking the enum non-exhaustive forced every
downstream match — mapping to a gRPC status, say — to carry a _ arm that can
never fire, and denied them the compile error if a variant ever is added.

Removing the attribute is backwards compatible; adding it back would not be, so
before 0.1.0 ships is the moment.

Query derives Default

Every field already had a Default, and three of the four are absent on a first
unfiltered page:

let query = Query {
    order_by: "name".parse()?,
    columns: VOLUME_COLUMNS,
    ..Default::default()
};

The default Columns map is empty, which rejects every path — the derived
default is the fail-closed one, not an open door.

The placeholder-style check has a name

is_positional, with the reasoning attached: that it infers a behavioural
property from rendered text, that this is a smell, and that the honest fix is a
Dialect::is_positional in sqlx-cel defaulting to exactly this comparison. I'll
raise that upstream separately.

I also walked back a claim I made about the ordinal arithmetic in the positional
branch: param_offset + repeated.len() - 1 is not dead, it is the true
parameter ordinal of the bind. A positional dialect ignores the number, but it
is the honest answer, and the comment now says so instead of the code changing.

The bug this found

Writing the helper, I inverted the comparison. != classifies every numbered
dialect as positional, which repeats binds that are already addressable and
shifts the entire page. Seven existing tests failed, which is a decent answer to
whether they were load-bearing.

The new test pins the case that motivated asking the dialect rather than
hardcoding the three built-ins: SQLite's ?1 form is positional in syntax but
still addressable, so its values must not repeat. A "does the placeholder
contain ?" check would get that wrong.

Testing

34 unit tests, 6 doctests, 6 SQLite round-trips, 6 Postgres round-trips; fmt,
clippy and cargo doc -D warnings clean across the feature matrix.