[avison9]

Closes #4183

Rationale for this change

A transaction that writes two refs only guarded the first one against concurrent changes.
Transaction._stage de-duplicated requirements by type alone, so after the first write staged
AssertRefSnapshotId(ref="b"), the second write's AssertRefSnapshotId(ref="main") was dropped and
the commit reached the catalog without an assertion on main.

The type-only de-duplication was correct when it was added (#569), because a transaction could only
ever assert main. Writing to branches (#941) made AssertRefSnapshotId per-ref, but the
de-duplication was never updated.

This changes the key to (type, ref) for AssertRefSnapshotId. Each ref keeps its first (base)
assertion, matching Java's UpdateRequirements. Other requirement types are unchanged. Retries go
through _stage too, so they now keep every ref's assertion as well.

cat.load_table("ns.t").manage_snapshots().rollback_to_snapshot(s1).commit()   # another writer
with t.transaction() as tx:
    tx.append(row(3), branch="b")
    tx.append(row(3))

cat.load_table("ns.t").scan()   # before: [1, 2, 3]  rollback silently undone
                                # after:  [1, 3]     conflict detected, commit retried

Measured on a transaction that writes branch b and then main:

Concurrent change to main	Format	Before	After
rollback	v1 / v2	rollback undone	retried, [1, 3]
append	v1	appended row lost	retried, [1, 2, 3, 4]
append	v2	non-retryable ValueError (sequence number)	retried, [1, 2, 3, 4]

Are these changes tested?

Yes. Two tests were added to tests/table/test_commit_retry.py, and each runs against every catalog
fixture:

• test_transaction_writing_two_branches_detects_concurrent_change_to_either covers the issue's
  scenario on format v1 and v2, with both a concurrent rollback and a concurrent append.
• test_transaction_stages_one_ref_assertion_per_ref checks that writing b and then main twice
  stages exactly one assertion per ref, each holding the base snapshot id.

Both fail when the _stage change is reverted. Unit suite: 4524 → 4530 passed, no new failures.
make lint is clean. make test-integration passes (1362 passed, 0 failed).

Are there any user-facing changes?

Yes, as a bug fix. Multi-ref transactions now detect concurrent changes to every ref they write, and
retry or fail with CommitFailedException instead of overwriting them. There is no API change.

[Fokko]

LGTM! 🙌 The per-ref requirement key is the right fix, and the two-branch test that fails on main is exactly the regression this needed. Thanks @avison9!