fix(qwen3-vl): CP-local vision-embed hook + AllGatherVisionEmbeddings.apply kwarg

[Zhichenzzz]

What this PR does

Make the bridge's Qwen3-VL model work when the training framework passes CP-pre-sharded packed inputs (Megatron's standard load-balanced "zigzag" context-parallel layout), plus one small API fix:

1. Native CP-local vision-embed selection — Qwen3VLModel.forward computes vision embeddings for the full sequence and scatters them into the vision-token positions of input_ids. When input_ids are already sharded across CP ranks (detected via cu_seqlens_q[-1] == cp_size * len(input_ids)), the local vision mask covers only this rank's vision tokens, so the rows must be filtered to match. Qwen3VLModel._cp_local_vision_embed_indices computes the selection natively: it all-gathers only per-chunk vision-token counts (a few ints per segment — not the token ids) and places this rank's vision tokens in the full vision-token order. Stock full-sequence behaviour is unchanged (the method returns None for full inputs, cp=1, or unrecognized layouts).
2. AllGatherVisionEmbeddings.apply kwarg fix — torch.autograd.Function.apply rejects keyword arguments; pass cp_group positionally at the two call sites.

Design notes

• An earlier revision exposed an overridable identity hook for frameworks to inject the selection; review feedback rightly flagged that as awkward, so the selection is now implemented natively (the zigzag layout is Megatron's own convention, not framework-specific).
• Control flow before the count all-gather only depends on cu_seqlens (identical on all CP ranks) — rank-dependent shortcuts happen strictly after the collective, otherwise the CP group deadlocks.
• A rank whose chunks contain no vision tokens selects zero rows (empty index), matching its empty local mask.

Validation

Qwen3-VL-2B, CP2 TP4 and CP2 TP2, THD packing, geo3k RL with miles (radixark/miles#1308): coherent rollouts, train_rollout_logprob_abs_diff 0.0127–0.0131 (same healthy band as non-CP runs), no NCCL timeouts, stock non-CP path unchanged.

Companion PR: radixark/miles#1308 (per-segment mRoPE positions for the same pre-sharded path; its vision plumbing is deleted in favour of this native support).