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 through resume() and restart an async loop only when it was previously running
• expose 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 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

Implementation commit: 5900c786322b522162af0bef0674464806f1628a.

A separate three-L20 stability soak ran 100 cycles per context with the tiny deterministic preset:

Max context	Cycles	Reclaimed	Release median / p99	Resume median / p99	Greedy parity
2,048	100	0.572 MiB	72.201 / 88.869 ms	0.410 / 0.553 ms	100/100
8,192	100	2.239 MiB	74.530 / 100.495 ms	0.581 / 6.293 ms	100/100
32,768	100	8.907 MiB	68.257 / 93.089 ms	0.441 / 0.557 ms	100/100

All 300/300 post-resume greedy outputs matched. The tiny-preset footprint is stability evidence and is not directly comparable to the AstrAI-1B memory table.

Validation

• local: Ruff format/import checks passed; 628 passed, 103 skipped
• NVIDIA L20 focused scheduler/engine/rollout suite: 71 passed
• 100-cycle result and five-cycle raw per-sample JSON are checked in under docs/benchmarks/
• InfraSWE 811bc775: comparison cell valid; 53 Draft/system-path tests passed
• diagnostic ProjectFit 92.29/100, BenchmarkTrust 97.40/100

InfraSWE scoring is diagnostic and unsealed. CUDA graphs were disabled to isolate scheduler-owned state. The source-mounted runtime used established extension fallbacks, so this PR makes a lifecycle/allocation claim rather than a kernel-throughput claim.