Parallelize the preprod snapshot odiff comparison phase. The single serial OdiffServer is replaced with a pool of N persistent servers (tunable via the new preprod.snapshots.odiff-worker-count option, default 4) driven by a ContextPropagatingThreadPoolExecutor. The per-batch fetch → compare → upload work is extracted into _process_batch, which checks a server out of a queue.Queue, runs the batch, uploads its diff masks, and returns partial results that are merged single-threaded in the main loop.

Why: for large builds the comparison diffs thousands of changed image pairs one at a time against one node process, taking 3+ minutes — long enough to hit the task's 300s processing_deadline (ProcessingDeadlineExceeded) or be interrupted by a deploy SIGTERM. A real build (~6,000 changed pairs across 431 batches) was timing out this way. Spreading the batches across N servers brings it comfortably under the deadline.

Notes for reviewers:

• Concurrency safety: each OdiffServer is only ever used by one thread (checked out of the pool); the objectstore Session is thread-safe for get/put; each batch's compare_images_batch uses its own tempdir, so there are no cross-thread filename collisions. All result accumulation happens in the main thread — workers mutate no shared state.
• A mid-batch failure is isolated: already-processed pairs are preserved and only the unprocessed ones are marked errored, so a single bad batch doesn't fail the whole comparison.
• Memory scales with N (each worker can hold a pair's decoded/padded images), which is why N is an FLAG_AUTOMATOR_MODIFIABLE option — it can be dialed down in prod without a deploy. MAX_DIFF_PIXELS still caps any single pair.
• Adds a compare_snapshots: odiff phase complete summary log (duration, throughput, slowest batch, worker count) for post-deploy diagnosis, and the first end-to-end tests for compare_snapshots (which previously had no task-level coverage).

This is the durable follow-up to #116708 (which stopped killed comparisons from getting stuck in PROCESSING); together they address the timeout end-to-end.