[Fix] Conductor SIGINT, piecewise KV capture, and batched AC rollout split

[merceod]

What does this PR do?

Three unrelated fixes found while validating #164:

• api_server: _shutdown_conductor_process called send_signal(SIGINT) on a multiprocessing SpawnProcess, that's subprocess.Popen API, so every shutdown raised an AttributeError (caught and logged) and the conductor was always SIGTERM'd without running its graceful path. Use os.kill instead. SIGINT-ing the api server now yields "Shutting down conductor..." and a clean tree exit.
• engine: the piecewise runner still instantiated BatchedCacheManager directly; it has been an ABC since [WIP] cosmos3: add generator model scaffold (+reasoner implementation) #121, so every uses_kv_cache=True piecewise capture failed at startup (warn-only, easy to miss) and vjepa2_ac's block_loop graphs were never captured. Route through create_cache_manager() like the other call sites. All four buckets (bs=1/2/4/8) capture again.
• vjepa2: VJepa2ACRolloutPredictorSubmodule.forward_batched sliced the output list instead of the tensor when splitting per request, so with B≥2 the first request got the whole batched tensor and the rest got an empty list. Cconcurrent AC rollouts then crashed (Per-request tensor has leading dim 2 != 1) or truncated. Slice rows like the sibling submodules do. New CPU-only regression test test_ac_rollout_batched.py fails on the old code.

How was it tested?

Verified end-to-end on vjepa2_ac: sequential outputs bit-identical to before the change (sha-matched), and 6 concurrent rollouts at concurrency 3 all complete where they previously crashed.

Checklist

• ruff check . passes
• Added or updated tests / docs where relevant

[NSagan271]

LGTM, also fixed a few other vjepa issues