Make the hackbot action apply path safe to repeat and hard to poison

[msujaws]

Split out of #6439 so it can be reviewed and landed on its own: these are fixes to
the existing apply path, not part of the frontend-triage auto-apply feature.

Both bugs are reachable today, on the manual apply-all path. They matter more once
anything applies actions unattended (#6439), which is why they were found — but
they stand alone and change no user-visible behaviour.

Concurrent deliveries could double-post

Pub/Sub push is at-least-once and concurrent, so two deliveries of the same
run.completed could both read a pending row and both post its comment to the bug.
A manual apply-all racing the automatic path could do the same.

Rows are now locked before dispatch and stay locked until the result is committed, so
the second caller blocks, then reads applied and skips. Holding the lock rather than
flagging the row in-progress is what keeps this from needing recovery: a process that
dies mid-dispatch rolls back to a row exactly as retryable as it was — no in-flight
state to reclaim, and no new column to migrate. The cost is a transaction open across
the Bugzilla call, which is why the lock is per coalesced group rather than per pass.

Four inputs could 5xx the push route in a loop

There is no dead-letter topic on this subscription, so a 5xx replays the same message
for the whole retention window. Each of these now fails a row a human can read instead:

• two concurrent first deliveries racing the row insert (now ON CONFLICT DO NOTHING)
• a recorded action with no type, or params that aren't a mapping
• a params shape that raises inside the coalescer
• an unresolved {{actions.<ref>.<field>}}, which was previously sent literally into a
  real bug comment with only a log line as the signal

Notes

• No schema change and no migration.
• No UI change.
• Adds 12 tests; the 3 failures and 8 errors in services/hackbot-api/tests are
  pre-existing on master (an is_markdown coalescing mismatch and a missing client
  fixture, in files this PR does not touch).
• Conflicts with Auto-apply high-confidence frontend-triage results #6439 by construction — both touch actions_applier.py. Whichever
  lands second needs a rebase.