perf(zero-cache): improve rm fanout flow control

[Karavil]

what changed

Raises the RM fanout byte window, batches forwarded changes, and lets flow control release after majority completion plus padding.

Also fixes the split branch ordering issue found by benchmark: a subscriber batch now enqueues the whole batch synchronously before awaiting results, so later fire-and-forget batches cannot interleave before a prior commit.

benchmark output

Targeted 16-subscriber websocket run, medium-wide-batch-pressure, 15s, apply disabled:

tx/s: 273.7 -> 306.4
rows/s: 5,474.3 -> 6,128.0
fanout msg/s: 84,481.0 -> 91,086.5
max ACK lag: 47,828 -> 47,498

The single-consumer stack run is roughly neutral to noisy for this PR, which matches the intent. This change is about RM fanout pressure across multiple serving replicas.

validation

pnpm exec vitest --project='*no-pg*' run src/services/change-streamer/broadcast.test.ts src/config/zero-config.test.ts src/services/change-streamer/subscriber.test.ts
pnpm run check-types
pnpm exec oxlint --config ../../oxlint.config.ts src/services/change-streamer/broadcast.ts src/services/change-streamer/forwarder.ts src/services/change-streamer/subscriber.ts

[Karavil]

Replaced by rocicorp/mono PRs rocicorp#6070 through rocicorp#6078.