Fix silent write-loss in mt_upsert_<doc> + ProjectLatest double-store regression

[jeremydmiller]

Supersedes #4325. Picks up @ArieGato's three commits unchanged and adds one
follow-up commit that fixes the regression those changes expose in
project_latest_tests.inline_projection_stores_document_on_project_latest.

What #4325 fixes

The generated mt_upsert_<doc> for documents with UseNumericRevisions = true
ran INSERT ... ON CONFLICT ... DO UPDATE ... WHERE revision > mt_version
and then unconditionally re-read mt_version. Under READ COMMITTED contention
the conditional UPDATE could match zero rows while the function still returned
the now-bumped mt_version, so Marten saw a non-zero return as success,
SaveChangesAsync did not throw, and AfterCommitAsync fired for a write that
never landed.

The fix folds the read into the same statement
(DO UPDATE SET ... RETURNING mt_version INTO final_version) and
RETURN COALESCE(final_version, 0) so that a filtered conflict update produces
0, which Marten already treats as ConcurrencyException.

Regression test in numeric_revisioning.cs (concurrent_update_revision_does_not_silently_lose_writes)
fails on master, passes here.

What this PR adds on top

FetchInlinedPlan.ProjectLatest was queueing an upsert at the projected
revision via session.Store(snapshot), and the configured inline projection
queued a second upsert at the same revision during SaveChangesAsync. Under
the old write-loss behaviour the second write silently became a no-op. With
the upsert function now reporting 0 when its WHERE filter skips the row,
the second upsert correctly surfaces as ConcurrencyException — and the
existing test inline_projection_stores_document_on_project_latest started
to fail in CI.

The inline projection persists the projected document itself during
SaveChangesAsync, so the explicit Store in the inline fetch plan was
always redundant; it just happened to be harmless under the buggy upsert
function. FetchAsyncPlan keeps its session.Store because async
projections don't run inline.

Test plan

Run on net10.0 against PostgreSQL 16:

• EventSourcingTests filter project_latest_tests — 6/6 pass (was 5 pass / 1 fail on the previous PR head)
• DocumentDbTests filter numeric_revisioning|Concurrency — 66/66 (includes the new race regression test from Fix silent write-loss in mt_upsert_<doc> for numeric-revisioned documents #4325)
• EventSourcingTests filter numeric_revisioning|Concurrency — 21/21
• EventSourcingTests filter Aggregation|Projection — 409/409
• EventSourcingTests filter FetchForWriting|FetchLatest|ProjectLatest|InlineProjection|inline_aggregation — 212/212
• EventSourcingTests filter Bug_3310_inline_projections|SubscriptionScopedSession|MultiStream — 22/22
• PatchingTests — 114/115 (1 pre-existing skip)
• MultiTenancyTests filter revisioning|Concurrency|Aggregation — 1/1
• CoreTests filter migrate_from_guid_to_int_based_revisions — 1/1

🤖 Generated with Claude Code