Move in-model CP layout helpers into miles_plugins/models/cp_utils.py

[yueming-yuan]

What

Pure refactor, no behavior change.

miles/backends/training_utils/cp_utils.py owns the CP helpers the training backend applies around the model — slice_with_cp, natural_to_zigzag_slice, slice_loss_masks_for_local_cp, all_gather_with_cp, allgather_cp_redistribute, get_logits_and_tokens_offset_with_cp, … All of them operate on tokens / masks / logprobs / logits outside forward(), and are called from loss.py, data.py, losses.py.

build_gdn_cp_context was the one exception. It runs inside a layer's forward, its only callers are miles_plugins/models/qwen3_next.py and qwen3_5.py, and it forced an optional flash-linear-attention dependency into a core training module:

try:
    from fla.ops.cp import build_cp_context as _fla_build_cp_context
except ImportError:
    _fla_build_cp_context = None

The zigzag ↔ contiguous-shard relayout has the same shape but the opposite problem: it lived as private helpers inside miles_plugins/models/hf_attention.py, so only models built on HuggingfaceAttention could reach it. It is not HF-attention-specific — it bridges Megatron's zigzag CP layout to the contiguous rank-local shard every fla CP operator expects, and any model with an fla linear-attention layer needs it.

Change

• miles_plugins/models/cp_utils.py becomes the module for CP helpers that run inside a forward: build_gdn_cp_context (moved from core) plus the relayout helpers zigzag_to_packed_shard / packed_shard_to_zigzag and their internals (moved from hf_attention.py, underscore prefix dropped).
• detect_and_setup_hybrid_cp moves the other way, out of miles_plugins/models/cp_utils.py into hf_attention.py. It is defined in terms of isinstance(module, HuggingfaceAttention), so it belongs next to that class — and moving it is what breaks the import cycle hf_attention → cp_utils → hf_attention.
• Core cp_utils.py loses build_gdn_cp_context, the fla.ops.cp try/except, and its now-unused torch.nn import.
• Import sites updated: qwen3_next.py, qwen3_5.py, miles/backends/megatron_utils/actor.py, tests/e2e/precision/test_hf_attention_cp_relayout.py.

Verification

• Every relocated definition is byte-identical to its original apart from the underscore renames — checked by extracting both versions with ast and comparing source text (10/10 identical).
• pre-commit run --all-files clean.
• Static module-level import-graph check: no cycles involving any touched module. miles_plugins.models.cp_utils now has zero intra-repo imports; core cp_utils imports only .parallel.
• Behavioral coverage is unchanged and comes from the existing tests/e2e/precision/test_hf_attention_cp_relayout.py round-trip test, which now imports the public names. It needs GPUs, so it runs in CI rather than locally.

Motivation

New model support adds a third model that needs the relayout, and it is not a HuggingfaceAttention subclass. Without this, it would either duplicate the permutation logic or import a private symbol across plugins. This PR stands on its own, though — it removes an optional-dependency leak from a core training module either way.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[Zhichenzzz]

LGTHuman