--allgather-cp gives wrong log-probs and NaN grads on stock Qwen3 attention #4469

--allgather-cp gives wrong log-probs on stock Qwen3, and the first training step fails with "found NaN in local grad norm". Seen on main (39b8e437ce), with the torch and fused log-prob backends.

Repro: Qwen3-1.7B, Megatron GRPO, --context-parallel-size 2 --allgather-cp, replaying rollouts recorded with sglang. The trainer's mean log-prob of the sampled tokens is about −0.82 (−0.48 with micro-batch size 1), against −0.296 from the rollout. The same run without CP (−0.2968) or with zigzag CP2 (−0.2967) matches the rollout and trains without NaN.

Likely cause (not verified line by line): --allgather-cp changes miles' data layout to one contiguous chunk of the sequence per CP rank, but only the DSA, DeepSeek-V4 and fla/hf_attention model paths read that layout. Megatron's stock attention for Qwen3 still assumes zigzag chunks, so its attention sees the wrong positions.

Suggested fix: reject --allgather-cp at startup for models whose attention doesn't support the contiguous layout, with a message naming the supported ones; or make the stock path support it.

Found while measuring #4228 (which is unaffected: the bug is the same with and without it).