feat: release and resume inference runtime

[0z5a]

Summary

• add a reversible lifecycle for scheduler-owned inference runtime state
• release KV pages, task caches, executor workspaces, and graph state while retaining model weights and policy version
• reconstruct the runtime through resume() and restart an async loop only when it was running before release
• expose the lifecycle through InferenceEngine, RolloutGenerator, and RolloutRunner, including rollout tensor-cache cleanup
• reject release for caller-owned caches instead of silently discarding external state

NVIDIA L20 benchmark

AstrAI 1B, BF16, batch 4, prompt 128, four greedy decode tokens, five complete release/resume cycles per cell:

Max context	Runtime footprint	Reclaimed	Reclaimed %	Release median	Resume median	Greedy parity
2,048	200.97 MiB	192.85 MiB	95.96%	105.62 ms	2.01 ms	5/5
8,192	777.14 MiB	769.01 MiB	98.95%	90.16 ms	5.09 ms	5/5
32,768	3,081.81 MiB	3,073.68 MiB	99.74%	89.27 ms	5.10 ms	5/5

The benchmark is bound to implementation commit 5900c786322b522162af0bef0674464806f1628a; raw per-cycle JSON and the reproduction tool are checked in. CUDA Graph was disabled to isolate scheduler-owned KV/workspace memory. The source-mounted image did not expose the optional attention, rotary, and GEMV extensions, so it used the established fallbacks; this PR makes a runtime-allocation lifecycle claim, not a kernel-throughput claim.

Validation

• local: Ruff format/import checks passed; 628 passed, 103 skipped
• NVIDIA L20 GPU5: focused scheduler/engine/rollout suite 71 passed in 2.27s
• InfraSWE commit 811bc775ed5b3a6ec853219245f3469f78818020: comparison cell validated and 53 Draft/system-path tests passed
• visible-evidence diagnostic ProjectFit: 92.29/100
• BenchmarkTrust: 97.40/100

The InfraSWE score is explicitly diagnostic and unsealed. Official ProjectFit remains unresolved because the Draft seal, system-trace evidence, hidden probes, and verified evidence manifest are absent.