[merceod]

Stacked on #1095 (pr-29-s5-pinned-arena).

Why

A cold server on MI325X compiles for minutes: r9b takes 340 s on Qwen3.5 9B, 644 s on Whisper and 1032 s on Qwen3-Omni colocated with empty caches (one start each, one GPU, shared node). A restart on the same host is warm, but a new host or container pays all of it again, and there was no way to carry the Inductor, Triton and MIOpen caches over.

What changed

• mstar/cli/cache.py, mstar/utils/compile_bundle.py: mstar cache key|dirs|pack|unpack|show. A bundle is keyed on product, torch, HIP or CUDA, and Triton, and holds one uncompressed tar per cache. pack appends under a lock and merges MIOpen's databases. unpack refuses another stack unless given --force, and exits 3 when the root has no match, so callers can serve cold.
• mstar/cli/main.py: mstar serve --compile-bundle ROOT unpacks before serving.
• mstar/worker/worker.py: each worker logs its stack key and whether the unpacked bundle matches.
• mstar/engine/accelerator_graph_config.py, accelerator_graph_runner.py: a compile_dynamic capture option, one dynamic-batch graph for every bucket. Default off, and nothing here opts in.
• docs/compile_cache.rst, docs/serving.rst.
• All of it is opt-in and works by stack key, so it isn't platform specific. Without the flag the only change on any platform is the startup line. It hasn't run on NVIDIA.

Verification

• CPU: ruff clean. test_compile_bundle.py, test_accelerator_graph_capture.py, test_platform_dispatch.py, test_torch_config.py 90 passed, 4 skipped. On the speech stack's tip, the targeted set plus every branch's tests 2436 passed, 61 skipped, 0 failed, none failing only here. The widened 96-file set 1549 passed, 206 skipped, 5 failed, r9b 1519, 204, 5, the same 5 (ZMQ tests over the socket path limit under a long --basetemp, 15 passed with a short one).
• GPU, one MI325X: test_accelerator_graph_capture 29 passed.
• No window times startup. Seconds from launch to /health, MI325X, one start each, one GPU on a shared node, empty Inductor, Triton and MIOpen dirs, the speech stack's tip against r9b:

config	r9b cold	tip cold	tip bundle, 1st	tip bundle, 2nd
qwen3omni_colocated	1032	746	391	250
qwen3tts_1p7b	202	226	51	39
whisper_large	644	551	433	97
qwen3_5_9b	340	325	262	148

• Pack 1.4-3.7 s, 2.9 GB with all four rows, unpack 2.8-3.8 s.
• On the branch's own base, medians of three starts, same kind of job: 9B 337 to 136 s, Omni 848 to 247, Whisper 63 to 33, FLUX.2 klein 1140 to 54. A 3.07 GiB bundle packs in 11 s and unpacks in 4 s.
• Output from a bundle start: bit-identical on Whisper and FLUX, within the near-tie protocol on 9B and Omni.

Not in this PR

• The first start from a bundle is slower than the next on every row (1.3-4.5x here) and grows the Triton dir. The cause is unverified, so treat the 2nd-start column as the good case, not a startup claim.
• Omni still spends 211 s warming before capture with every cache warm. Re-tracing per batch-size graph is a hypothesis.
• Whisper's r9b cold start is 644 s because Whisper/Qwen3-ASR PR 5: serving performance #568 compiles one encoder graph per batch bucket. A dynamic-batch encoder on compile_dynamic compiles 4x faster, but its replay cost is unsettled and it stays on a local branch.
• The bundle format needs to go upstream with the kernel registry.