fix(qwen-image): drop the txt_seq_lens kwarg removed by diffusers

[Rockdu]

What

Stop passing txt_seq_lens into QwenImageTransformer2DModel.forward. It stays a
per-sample input — collate_cond_for_sample_batch still reads it to derive the text
mask and the padding width — it just no longer travels in the collated dict that the
trainer splats into the model.

Why

txt_seq_lens was removed from both QwenImageTransformer2DModel.forward and
QwenEmbedRope.forward after diffusers 0.39. Under the pin in #189
(f53d5520, post-0.39 main) the qwen-image training forward raises
TypeError: forward() got an unexpected keyword argument 'txt_seq_lens'.

Numerically a no-op on 0.38

Removing the kwarg cannot move a single bit on the current pin:

• In 0.38 the parameter only triggered a deprecate() warning in
  QwenImageTransformer2DModel.forward — nothing read it.
• The RoPE length comes from compute_text_seq_len_from_mask (identical in 0.38 and
  f53d5520), passed to pos_embed as max_txt_seq_len=text_seq_len. That path never
  consults txt_seq_lens.
• QwenEmbedRope.forward would fall back to max(txt_seq_lens) only when
  max_txt_seq_len is None, which the transformer never does.

Verified bitwise under #189's diffusers pin

tests/e2e/short/test_qwenimage_pickscore_grpo_5xGPU.py run on this branch inside the
pinned-diffusers image, compared against the standard recorded on 0.38:

image	radixark/miles_diffusion:test-pr-189
diffusers	0.40.0.dev0 @ f53d552036a0d1bd5570782a39cd40cfabf112bc
sglang	7605529
hardware	5x H200
command	--num-rollout 2 (the test's own args, unmodified)

metric series	points	vs standard
rollout/reward/raw_num_samples	2	identical
rollout/reward/raw_mean	2	identical
rollout/reward/raw_median	2	identical
rollout/reward/raw_std	2	identical
train/log_prob_old_idx_0	4	identical
train/log_prob_new_idx_0	4	identical
train/log_prob_mean_abs_diff	4	identical
train/model_output_mean_abs_diff	4	identical
train/model_output_rel_max	4	identical
train/grad_norm	4	identical

[e2e-metrics] PASSED: 10 metric series match test_qwenimage_pickscore_grpo_5xGPU.json

32/32 points bit-for-bit, so the 311 commits between v0.38.0 and f53d5520 leave the
qwen-image train and rollout numerics untouched. The e2e is the guard for this change,
hence no new unit test.

Not covered here

test_wan22_pickscore_grpo_17xGPU_single_node_4xGPU_proxy still OOMs under #189's pin.
That failure is memory headroom, not numerics: at the OOM the trainer needed 1.31 GiB
with 1.25 GiB free while a rollout engine still held 25.3 GiB and the colocated
pickscore reward 4.7 GiB on the same device. The wan-side diffusers delta is a fp32
rope buffer pin (~0.5 MiB) and one added hidden_states.contiguous() (~32 MiB after
the ulysses-4 split) — enough to tip a device already at 99% occupancy, and tracked
separately.

[zhihengy]

LGTM :)