fix: return upsert identity keys and correlate on the full key set so bulk upserts don't drop records

[DGollings]

Problem

Since #786, bulk upserts can silently drop records from the returned result while still
writing them to the database — so after_action hooks and notifiers never run for the
dropped records. This surfaced on the 2.10.0 → 2.11.0 upgrade (same regression as #808/#809).

bulk_create/3 correlates each returned row back to its changeset by the upsert identity's
keys. Two things break that correlation:

1. The identity keys aren't selected. When the action's select excludes them
   (e.g. upsert_fields [] with select: [], as Ash.Reactor's bulk_create step passes),
   the returned rows don't contain the identity columns, so nothing correlates and every
   record is dropped. This is the fix from @marcnnn's fix: return upsert identity keys so bulk upserts don't drop records #808, with the regression tests from test: cover bulk upsert dropping records when select excludes identity keys #809.

2. Unset (nil) identity keys are missing on the changeset side. For a multi-field identity
   that leaves a key unset — a nullable field left nil, common with nils_distinct?: false
   natural keys such as [barcode, timestamp, order_id, route_id] — the changeset key is built
   with Map.take(changeset.attributes, keys), which drops the unset key. The changeset's
   correlation key then has fewer fields than the returned row's key (whose Map.take(row, keys)
   includes the key as nil), so it still never matches even after (1).

Fix

• Add the identity's attribute keys to the upsert RETURNING list so returned rows always
  carry them. Ash.Actions.Helpers.select/2 masks the extra fields back out of the records,
  so this doesn't widen what callers observe. Identity keys that aren't attributes (e.g. an
  identity over a calculation) have no column and are left out.
• Build the per-changeset correlation key over the full identity key set, defaulting unset
  keys to nil, so both sides have the same shape.

Tests

• bulk upsert returns inserted records when select excludes the identity keys (from test: cover bulk upsert dropping records when select excludes identity keys #809)
• bulk upsert returns updated records when select excludes the identity keys (from test: cover bulk upsert dropping records when select excludes identity keys #809)
• bulk upsert returns records when a multi-field identity leaves a key unset (nil) (new;
  covers facet 2)

The fix in (1) and the first two tests are originally by @marcnnn (#808, #809); this PR carries
them and adds the multi-field / nil-identity correlation fix and test on top.

Note for reviewers

On PostgreSQL 17 (MERGE path) the pre-existing test returns a skipped upsert whose identity contains nil (added by #794) fails with a unique-constraint error — that's a separate
nil-identity issue in the return_skipped_upsert? path, unrelated to and not addressed by
this PR.

Disclosure

Drafted with AI assistance, but every change was reviewed and validated: the failure is
reproduced by a failing test first, the fix is verified against those tests, and the whole
change was exercised end-to-end against a real 8k-test downstream suite on PostgreSQL 17.

---

• Bug fixes include regression tests

[zachdaniel]

🚀 Thank you for your contribution! 🚀