[Fix] Align Qwen3.5 tests with Transformers 5.14.1

[jayhenry]

Summary

This PR combines the minimal collection workaround from #2008 with the Transformers 5.14.1 compatibility fixes diagnosed in #2002, without migrating the wider test suite away from third-party parametrize.

• avoid binding the unused Transformers lazy module in test_qwen3_tile_embedding.py
• call the current Qwen3.5 dense decoder with keyword-only arguments
• preserve the outer vision boundary around video frame placeholders
• update Qwen3.5 chat-template expectations for the outer video boundary
• use the shared Transformers 5.14 vision interpolation helper with FP32 accumulation and a model-dtype result
• update Qwen3.5 MTP baselines for the current token sequence and fused kernels
• keep the existing text tolerance and apply the diagnosed tolerance adjustment only to the long image path

Root causes

Collection failure

Third-party parametrize==0.1.1 recursively scans module objects in decorator-frame globals. The top-level transformers lazy module can populate its dictionary during that scan, causing RuntimeError: dictionary changed size during iteration.

The module import in test_qwen3_tile_embedding.py was unused. Importing only the concrete AutoTokenizer symbol avoids exposing the lazy module to this scan while retaining the current parameterization framework.

Transformers 5.14.1 compatibility

• decoder arguments used by the parity test are now keyword-only
• current Qwen3/Qwen3.5 processors retain an outer vision boundary when expanding a video into frame placeholders
• positional interpolation moved to get_vision_bilinear_indices_and_weights
• the current fused GatedDeltaNet/MoE path introduces bounded numerical drift on the longer image sequence; instrumentation verified bitwise-equal vision output and language-model input before drift accumulates through the language layers
• chat-template regression strings must include the same outer video boundary

Validation

Using pt29_glm2, variables from zdev/env.sh, and gpu_lock.sh for GPU tests:

• five model/dataset files: 41 tests collected
• Qwen3.5 chat-template file: 5 passed
• Qwen3.5 dense decoder cases plus vision-tower bitwise parity: 3 passed
• Qwen3.5 standard VL, SP=1/4: 2 passed
• Qwen3.5 MTP, SP=1/4: 2 passed
• static audit: no test combines third-party parametrize with module-level import transformers
• production-file Ruff checks and git diff --check: passed

The previous Action completed with 558 passed, 18 skipped, 1 failed; the sole failure was the stale chat-template expectation fixed by commit 9be88872.

The full local suite was intentionally not run. Full HF video-reference cases require the CI TorchCodec/FFmpeg environment and are covered by the triggered Action.