Use stm-queue's bounded queues, add ActorConfig, monitors, and awaitEffects

[SamuelSchlesinger]

Summary

Two commits, releasing as 0.4.0.0 (Hackage has 0.3.1.1; 0.4.0.0 was never published, so the changelog entry is dated rather than bumped further).

1. Bounded mailboxes delegate to stm-queue (commit 1)

Bounded mailboxes were an unbounded stm-queue plus a single occupancy TVar that every send and every receive wrote. stm-queue-0.2.2.0 (stm-queue#3) provides bounded queues with split read/write credits, so capacity accounting conflicts once per capacity sends instead of on every message. The in-house Mailbox record and counter are deleted; semantics are unchanged and covered by the existing tests plus a new 8-sender × 500-message fan-in test.

2. API improvements from the review (commit 2)

• actWith + ActorConfig (mailboxCapacity, onCompletion, onUndelivered, onEffectFailure), with defaultActorConfig. act/actBounded/actFinally/actFinallyBounded are specialisations, unchanged in signature.
• Dead letters: messages still queued when an actor stops go to onUndelivered in mailbox order (only called when non-empty). Every committed send now reaches either a handler or onUndelivered.
• Effect failures: each throwing completion effect is passed to onEffectFailure; the default rethrows, preserving the old "first exception rethrown in the terminating thread" behaviour, but a logging handler can now replace stderr output.
• addAfterEffect is lifecycle-checked by default, matching send. addAfterEffectUnchecked keeps the old behaviour; addAfterEffectChecked stays as an alias.
• monitor / monitorSTM: message-based completion notification (Erlang 'DOWN'-style) alongside exception-based link. Immediate delivery for an already-stopped target, transactional with the registration.
• awaitEffects: waits until the completion handler and after-effects have finished (set in a finally, so it fires even if effects throw).
• murder is a no-op once the actor has stopped, so cleanup is never interrupted. (A murder racing the transition is caught by the effect runner and recorded as an effect failure — cleanup was never truncated, contrary to what the review first suggested.)
• GC behaviour documented and tested: a receiver whose handles are all dropped is stopped with BlockedIndefinitelyOnSTM, cascading through links/monitors.
• Benchmark stm-actor-benchmark [senders...]: fan-in throughput, unbounded vs bounded mailbox. CI now builds it under -Werror.

Effect order at shutdown: lifecycle transition → links/monitors initiated → onCompletion → onUndelivered → user after-effects → effectsFinished set.

Release ordering

stm-queue-0.2.2.0 is not on Hackage yet, so cabal.project pins it to commit b1c8f56. The main CI jobs pass with the pin; the sdist-outside-project and oldest-dependencies jobs will fail until it is published — that is what they are for. Order: merge stm-queue#3 → publish stm-queue 0.2.2.0 → drop the pin → merge this.

Test plan

• cabal build --enable-tests --enable-benchmarks all --ghc-options=-Werror (pinned project and a local checkout)
• cabal test all — 46 examples, 0 failures; 6 consecutive runs across -N2/-N4, including the GC test
• cabal check and cabal haddock --haddock-for-hackage clean
• Benchmark runs; bounded figures under fan-in are dominated by STM wakeup storms and are sensitive to -N (noted in the source)