SS-322: prune keys referencing excluded columns during PG purification

[patrickwwbutler]

Summary

• Postgres source purification pruned an excluded column from the stored table descriptor but left any key/constraint naming that column untouched.
• If the excluded column (and its constraint) was later dropped upstream, determine_compatibility's keys check saw the stale key vanish from the upstream key set and halted the source with "incompatible schema change", even though the column was never ingested.
• Fix: when pruning excluded columns in purify_source_exports, also drop any key whose columns reference an excluded column, so the descriptor doesn't carry a key that can never be re-verified once the excluded column disappears upstream.

Test plan

• Added a regression test to test/pg-cdc/exclude-columns-key.td covering both verify_schema call sites (replication and snapshot) for a table that excludes its primary key column and later has that column dropped upstream.
• bin/mzcompose --find pg-cdc run cdc exclude-columns-key.td passes, including the post-fix Materialize restart sanity check.
• cargo check -p mz-sql and bin/fmt are clean.

🤖 Generated with Claude Code

[martykulma]

I would add a couple tests, but otherwise looks good!

• composite primary keys (PRIMARY KEY (A,B)) with A excluded
• multiple keys (PRIMARY KEY (A), UNIQUE (B)) with A excluded

[patrickwwbutler]

Added both test cases: a composite PRIMARY KEY (id, sub_id) with id excluded (verifies the whole key is dropped, not left partially referencing the pruned column), and a table with independent PRIMARY KEY (id) + UNIQUE (u) with id excluded (verifies only the key naming id is dropped, leaving UNIQUE (u) intact).