Fix prefixed migrations: schema-qualified index names + cross-schema existence checks

[mdon]

Summary

An external user running the full migration chain into a named schema (--prefix "companyplexus") hit a hard failure. Investigating it surfaced two bug families in the versioned migrations, both invisible on the default public path:

1. Schema-qualified index names on CREATE INDEX (hard failure). Postgres rejects CREATE INDEX schema.name ... outright — an index always lands in its table's schema, so only the table may be qualified. (DROP INDEX schema.nameis valid, which is why the drop helpers using the same idiom never failed — those are untouched.) With a non-public prefix the chain died with syntax error at or near ".". Five sites fixed: ensure_fk_target_unique_indexes/2 and create_uuid_fk_index/4 in uuid_fk_columns.ex, add_uuid_unique_indexes in V56 and V57, and the V95 media-folders unique index. Note these errors surface at a later version's flush() (V61), not the version that queued the SQL.

2. Cross-schema-leaky idempotency checks (silent skips). Existence checks that don't anchor to a schema see objects from every schema. When a prefixed install runs in a database that also carries a public install, the check finds public's object and silently skips creating the prefixed one — the prefixed schema ends up missing columns/constraints (or hard-fails downstream, which is how the new test caught it). Fixed: V95's information_schema.columns check gained a table_schema filter, and 13 pg_constraintconname checks gained the AND conrelid = '<prefixed table>'::regclass anchor (V51's existing idiom): V35, V102 ×5, V113 ×2, V115, V118 up+down, V119 up+down. The other 43 pg_constraint sites were already anchored.

3. Regression test.test/integration/prefix_migration_test.exs runs the full 142-version chain into a scratch schema on the test database and asserts the version marker, the once-buggy indexes' placement, and a full index complement. Because the test DB also carries a public install, it fails on either bug family. It flips the sandbox to :auto for the run (the same boot path test_helper.exs uses) — a sandbox checkout can't host Ecto.Migrator's own runner connections, and a dynamic non-sandbox repo instance doesn't work either because V08's backfill resolves the repo by module name (Ecto.Adapters.SQL.query(RepoHelper.repo(), ...)), bypassing put_dynamic_repo/1. Details in the test moduledoc.

Verification

• Fresh scratch DB, prefix: "companyplexus": chain previously died at V61; now completes with all 730 phoenix_kit indexes in the prefixed schema, none leaked to public
• Fresh scratch DB, default public: completes (no regression)
• Prefixed chain into a DB with an existing public install (the new test): passes in ~5s
• Permissions/migration/smoke test files: 159 tests, 0 failures
• mix precommit (format + compile + credo --strict + dialyzer): green
• Reviewed by Codex: confirmed no missed sites of either family, ::regclass casts safe (every target table is guaranteed to exist at its check), and prefixed-vs-public semantics correct; its one finding (test cleanup could leave the sandbox in :auto if the schema drop raises) is fixed with try/after