feat(orchestrator): route per-queue storage through host profiles and fix the buildsignal republish dedup - #538

… fix the buildsignal republish dedup

## Summary

### Why?

Per-queue backend routing is host policy: profiles.go already routes every other seam (build runner, change provider, analyzer) per queue, so storage joins the same registry now that its contract is factory-resolved. Separately, end-to-end validation caught a regression from the payload change: repointing buildsignal's speculate publish from the batch-ID partition to the queue partition made its message share the (topic, partition, message-id) triple with the batch controller's original speculate publish, and the MySQL queue backend dedupes publishes on that triple against rows not yet garbage-collected — so the "build finished, re-evaluate" nudge was silently dropped and the batch stalled in Speculating.

### What?

`Profile` gains a per-queue storage factory entry; every profile points at the shared MySQL backend by default and `Profiles.StorageFactory()` routes each resolution through the queue's profile, so a deployment that splits queues across backends overrides one field. The orchestrator's `Deps.Storage` now comes from the profile registry. Buildsignal keeps the queue partition (preserving per-queue serial ordering) but mints its speculate message ID from (batch, build) instead of reusing the bare batch ID, so the re-evaluation publish never collides with the batch controller's original message while a redelivered poll still dedupes against itself.

## Test Plan

✅ `make e2e-test` (submitqueue suite green — reproduces the Speculating stall without the message-ID fix) ✅ gateway + orchestrator integration suites via bazel ✅ `go test ./...` ✅ `make fmt` ✅ `make lint` ✅ `make check-tidy` ✅ `make check-gazelle`. Local docker-under-bazel runs need `--sandbox_writable_path=$HOME/.docker` for the image builds.