refactor(fsdp)(4.3/5): family-declared input dtype policy at the model boundary

[Rockdu]

Rebased onto main now that #100 (refactor 4.2) is merged. Weight/param dtype control is not in this PR: the per-parameter dtype patch landed via #100; this is the input side only.

---

What

• Each model family declares the dtypes of its model-boundary inputs (latents, cond, timestep) as an input_dtype_policy on its TrainPipelineConfig: a dtype name, "default" (the run's forward dtype), or None = pass the rollout dtype through. Default policy is all-None.
• The trainer stops hard-casting forward inputs (latents/timesteps .to(forward_dtype) and cast_cond_to_dtype removed; FSDP cast_forward_inputs=False) and instead applies the family policy once at the boundary, then runs the DiT forward under torch.autocast(forward_dtype).
• LTX opts in (latents/cond = "default", timestep passthrough) and its inner per-forward autocast folds into the trainer's ambient one.

Why

Boundary input dtypes are what element-wise ops see before any weight is involved, so they must match what the family's sglang-d pipeline feeds the DiT for train/rollout log-prob alignment — and that is per-family knowledge, not a global cast. With the boundary owned by the policy, compute is owned by autocast; the grad-ckpt buffer force-cast in apply_fsdp2 is dropped because autocast keeps recompute dtypes consistent without rewriting fp32 buffers (e.g. RoPE tables).

Files

• miles/backends/fsdp_utils/input_dtype_policy.py — the policy keys and apply_input_dtype_policy
• miles/backends/fsdp_utils/actor.py — boundary policy application + ambient autocast; cast_forward_inputs=False; buffer force-cast removed
• miles/backends/fsdp_utils/configs/train_pipeline_config.py — input_dtype_policy field (default passthrough)
• miles/backends/fsdp_utils/configs/ltx.py — LTX policy; inner autocast removed
• miles/backends/fsdp_utils/loss_hub/{flow_grpo,nft,utils}.py — cond-cast removal (utils.py deleted, it only held the cast)
• tests/fast/backends/fsdp_utils/test_input_dtype_policy.py — 4 CPU tests (passthrough default, family override, float-only casting, unknown-key rejection), registered to stage-a-cpu
• tests/ci/fixtures/e2e_standards/test_sd3_ocr_grpo_2xGPU.json — re-recorded: autocast + boundary-cast changes shift SD3 numerics

Checklist

• pre-commit run --all-files passes
• Added/updated tests for new behaviour
• pytest -x is green — not run locally (no local torch); the same 4 policy tests were green on this PR's previous CI run, and stage-a-cpu re-runs them here
• If launch flags changed, python3 train.py --help still parses — no flags changed
• If a public flag was added, it appears in the CLI reference docs — no flag added
• If an example was added, it has a real walkthrough — no example added