[Feature] Enhance hadamard transform

[chensuyue]

Status:

1. support deterministic_hadamard_matrix and deterministic_hadamard_matrix
2. support inference with transformers with a triton mxfp4 kernel
3. support rtn and autoroud tuning (iters>0)

Todo:

1. align the implementation of SpinQuant: LLM Quantization with Learned Rotations, especially R2/R3, and implement structured fused
2. support online transform for vllm inference
3. support nvfp4
4. verify multi-cards

Originally posted by @lkk12014402 in #1323

[hshen14]

#1631

[lkk12014402]

target 0.13.0

1. fix [Bug]: hadamard dtype and inplace tranform #1631
2. support quip-style transform
3. align the implementation of SpinQuant: LLM Quantization with Learned Rotations, especially R2/R3, and implement structured fused
4. support online transform for vllm inference (mxfp4) with vllm_ext
5. support nvfp4/mxint4
6. compare turboquant/polarquant transform implementation

[lkk12014402]

fix hadamard transform dtype #1665