refactor(fsdp): sleep/wake module, CPU-offloaded models stay put, frozen models reuse host copies

[Rockdu]

Stack 3/9 of follow-ups to #263: main ← #265 ← #266 ← this ← #268 ← #269 ← #271 ← #272 ← #273 ← #274.

What

• refactor(fsdp): leave natively CPU-offloaded models out of sleep and wake: as in miles, a model under FSDP's CPUOffloadPolicy never sleeps; its parameter shards stay on the host and its buffers on the GPU. This drops the keep_parameters_on_cpu special case.
• refactor(fsdp): replace offload_model and onload_model with move_model in sleep_wake.py: one move_model(model, device) next to move_optimizer, in a module named after the trainer's sleep/wake phases (offload/onload stays the rollout's vocabulary). Rename only.
• perf(fsdp): reuse frozen models' pinned host copies ...: frozen reference and teacher models get wake_up_frozen_model / sleep_frozen_model. Waking keeps each parameter's pinned host copy, and sleeping points the parameters back at it instead of copying unchanged weights device-to-host again. Buffers are still copied, since a forward may update them.

Why

Reference and teacher weights never change after load, so the per-sleep device-to-host copy is pure overhead. Pinned host memory should not grow, because the caching host allocator already kept those blocks between phases (not measured).

Validation

OCR GRPO benchmark (SD3.5-M LoRA, KL 0.04 via lora_base, 2×H200, 6 rollouts, steady-state means excluding the first step), main vs #263:

	main	#263
actor_train	70.20 s	70.34 s
sleep	2.77 s	1.55 s
wake_up	0.94 s	0.22 s
sglang release (steady state)	0.53 s	0.53 s

The per-microbatch lora_base reshard costs ~0.14 s per step and sglang release is unchanged, so neither is changed here. That recipe has no independent reference or teacher; this PR's path is covered by the GPU offload worker, which checks that a frozen model's second sleep keeps the same host storage.

Files

• miles/backends/fsdp_utils/sleep_wake.py (was offload.py): move_model, move_optimizer; wake_up_frozen_model / sleep_frozen_model in their own section.
• miles/backends/fsdp_utils/actor.py, model_loader.py: which models sleep; frozen models sleep and wake through the frozen functions, each with its own frozen_parameter_host_copies.
• tests/fast-gpu/backends/fsdp_utils/{test_sleep_wake,_sleep_wake_worker}.py (were test_offload / _offload_worker): natively CPU-offloaded models only move their optimizer and EMA; a frozen model's sleep keeps each evenly sharded parameter's host storage.
• docs/fsdp_model_roles.md: sleep/wake wording and test paths.

Checklist

• pre-commit run --all-files passes — not run locally
• Added/updated tests for new behaviour
• pytest -x is green — GPU worker runs in CI only
• If launch flags changed, python3 train.py --help still parses — n/a
• If a public flag was added, it appears in the CLI reference docs — n/a
• If an example was added, it has a real walkthrough — n/a

[zhihengy]

LGTM