CT cast_to_fp4 torch recompile issues

[kylesayrs]

When running basic generation with NVFP4 models, I sometimes see torch recompilation issues.

Occurred while running examples/quantization_w4a4_fp4/qwen_30b_a3b.py

========== SAMPLE GENERATION ==============
[transformers] The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.
W0612 01:46:22.922000 871313 torch/_dynamo/convert_frame.py:1853] [0/8] torch._dynamo hit config.recompile_limit (8)
W0612 01:46:22.922000 871313 torch/_dynamo/convert_frame.py:1853] [0/8]    function: 'cast_to_fp4' (/home/kylesayrs/compressed-tensors/src/compressed_tensors/quantization/quant_args.py:55)
W0612 01:46:22.922000 871313 torch/_dynamo/convert_frame.py:1853] [0/8]    last reason: 0/7: tensor 'x' rank mismatch. expected 4, actual 3
W0612 01:46:22.922000 871313 torch/_dynamo/convert_frame.py:1853] [0/8] To log all recompilation reasons, use TORCH_LOGS="recompiles".
W0612 01:46:22.922000 871313 torch/_dynamo/convert_frame.py:1853] [0/8] To diagnose recompilation issues, see https://docs.pytorch.org/docs/main/user_guide/torch_compiler/compile/programming_model.recompilation.html

[Yatimai]

Root cause is the @torch.compile guard on input rank. During generation the model (a MoE) sends rank-varying tensors into cast_to_fp4, and since there is no automatic dynamic rank, each rank compiles its own graph until recompile_limit (8) is hit and dynamo falls back to eager. That's the rank mismatch warnings.

Fix flattens to rank-1 in an eager wrapper, keeping the compiled core from #331. One subtlety: flatten() works, reshape(-1) doesn't, because reshape(-1) infers a rank-dependent numel that defeats automatic dynamic shapes. Reproduced and verified on CPU (bug is rank-only, no GPU needed): graphs go from 5 to bounded across ranks 1-5. PR incoming.