feat: version rollout weight updates

[0z5a]

Summary

• track a monotonic policy version from optimizer steps through the inference scheduler and rollout results
• serialize synchronous generation with weight acknowledgements and invalidate reusable prefix KV pages on version changes
• preserve the behavior-policy version on cached rollout samples and initialize it correctly when resuming training
• document the weight-update and rollout attribution lifecycle

Control-plane benchmark

Exact PR commit 09bdc83, Python 3.12.2 and PyTorch 2.11.0+cu128 on the L20 server. Timings isolate InferenceScheduler.update_weights() after the scheduler is stopped and queues are empty, as required by the API.

version transition	state	median	p99 / range
monotonic acknowledgement	no cached prefix pages, 10,000 updates	1.48 us	p99 2.41 us
acknowledgement + invalidation	131,072 cached tokens, page size 64, 2,048 pages	2.43 ms	2.30–2.50 ms

The normal no-cache acknowledgement is microsecond-scale. Clearing a fully cached 2,048-page radix/allocator state takes about 2.4 ms and occurs only at the explicit weight-update barrier; it is not added to the decode hot path.

Validation

• PYTHONPATH=$PWD CUDA_VISIBLE_DEVICES=5 bash scripts/pre_commit.sh --skip-deps
• 643 tests passed
• ruff format and import checks passed

The full GPU suite used the existing compiled CUDA extension modules from the L20 checkout, linked into the isolated test worktree.

[ViperEkura]

landed as e58a728