Fix logits_to_keep support in Qwen2-VL forwards

[Terminator666666]

Summary

Fix logits_to_keep support in the Liger forward replacements for Qwen2-VL and Qwen2.5-VL.

The argument was previously captured by **kwargs, forwarded to the base multimodal model, and ignored when computing logits. The patched forwards now explicitly consume it and select the requested hidden-state positions before lm_head, matching Transformers 5.8.1 behavior.

Details

• Support both integer token counts and tensor indices.
• Preserve the default logits_to_keep=0 all-token behavior.
• Keep the fused-loss path on the complete hidden-state tensor.
• Add focused CPU tests for both model families, including argument forwarding and fused-loss coverage.

Testing Done

• Hardware Type: NVIDIA GeForce RTX 5070 Laptop GPU (tests used CPU tensors only)
• run make test to ensure correctness
• run make checkstyle to ensure code style
• run make test-convergence to ensure convergence

Targeted tests:

python -m pytest test/transformers/test_qwen2_vl_forward.py -q --confcutdir=test/transformers -p no:cacheprovider
11 passed

python -m pytest test/transformers/test_monkey_patch.py -k "qwen2_vl or qwen2_5_vl" -q -p no:cacheprovider
6 passed, 56 deselected

make checkstyle
All checks passed
317 files left unchanged

The full test and convergence suites were not run locally because this is a forward-compatibility fix validated with small CPU tensors; upstream CI can provide the broader GPU coverage.

[Terminator666666]

????(WSL Ubuntu,RTX 5070 Laptop GPU,BF16,?? test/convergence/bf16/test_mini_models_with_logits.py):

• mini_qwen2_5_vl:?? 32 ? + Liger 32 ?,loss?top-k logprobs?????????,1 passed,28 ??
• mini_qwen2_vl:?? 32 ? + Liger 32 ?,loss?top-k logprobs?????????,1 passed,29 ??

????????? tiny Shakespeare(batch size 16,sequence length 128),??????????

?????:11 ? CPU forward ?????6 ? monkey-patch ?????Ruff/checkstyle/py_compile/git diff --check ????