fix(qwen3.5-vl): cp_size 1 needs no packed-sequence divisibility check

[shifangx]

What

get_packed_cp_local_indices (slime_plugins/models/qwen3_5_vl_utils.py) rejects any
packed sequence whose length is not divisible by 2 * cp_size �~@~T including at
cp_size == 1, where there is nothing to split.

Reproduction

examples/geo3k_vlm/run_geo3k_qwen35.sh sets --context-parallel-size 1. Run as
written on GEO3K (Qwen3.5-35B-A3B, Megatron backend), it dies at the first
optimizer step of rollout 0:

slime_plugins/models/qwen3_5_vl.py:227       forward
slime_plugins/models/qwen3_5_vl.py:163       _inject_vision_embeddings
slime_plugins/models/qwen3_5_vl_utils.py:22  get_packed_cp_local_indices
ValueError: Packed sequence length 4551 must be divisible by 2 * CP size 1

Reproduced on 1 node �W 8 B200 and on 2 nodes �W 8 B200, failing on the same packed
length. Roughly half of all packed sequences have an odd token count, so this is
not a rare edge case: the Megatron VL path cannot complete a single gradient step
at CP 1.

Why

The function maps a CP rank's local tokens back to the packed stream using
Megatron's two-chunk zigzag layout, which splits every sequence into 2 * cp_size
chunks �~@~T hence the divisibility requirement. The check runs unconditionally, but at
cp_size == 1 one rank owns the whole stream and the correct local view is the
identity.

The fix

Short-circuit cp_size == 1 to the identity range and leave the zigzag path
untouched.

It changes no output that was previously produced: at cp_size == 1 with an even
length the existing loop already computes exactly that range (chunk = len / 2,
rank 0 takes [start, start + chunk) then [start + chunk, end)). It only stops
rejecting inputs that were always valid. Empty cu_seqlens now returns an empty
tensor instead of raising IndexError.

Tests

tests/test_qwen3_5_vl_native.py covered cp_size=2 with length-8 sequences �~@~T the
case that works. This adds test_thd_cp_indices_are_the_identity_without_context_parallelism
for the CP-1 odd-length and multi-sequence cases. The existing cp_size=2
expectations are unchanged.