[CUDA] columnwise quantize with tma

[nastya236]

Columnwise quantization with tma (mxfp8), bfloat16:

Size	With tma (ms)	Without tma (ms)
4096×4096	68.48	77.21
4096×8192	78.74	102.58
8192×4096	80.73	100.61
8192×8192	100.67	145.16
4096×16384	97.08	144.45
16384×4096	99.50	137.13

This PR:

• Adds PTX instructions for asynchronous copy with TMA
• Adds fp_quantize_columnwise_mxfp8 kernel for columnwise MXFP8 quantization using TMA on SM100+
• Splits fp_quantize.cu into fp_quantize.cu (dispatch) and fp_quantize.cuh (kernels) to reduce file size
• Moves swizzle_scales constants and get_swizzle_launch_args into cu:: namespace for consistency

The diff seems bigger than it is: since fp_quantize.cu already contains all the logic for block quantization, I decided to split it into fp_quantize.cu and fp_quantize.cuh to separate dispatch from kernel implementation.

TODO:
nvfp4 requires a separate columnwise kernel due to TMA tile size constraints. In the proposed kernel each thread processing a tile of size (N, M) and store a transposed result. M is equal to group_size -- 32 bytes formcfp8, but only 8 bytes for nvfp4. Since TMA requires the innermost tile dimension to be at least 128 bits (16 bytes), for nvfp4 kernel would need to load a larger tile and iterate over multiple groups.

[zcbenz]

Should we use the cuda::ptx APIs like cuda::ptx::mbarrier_init API instead? They don't have good documentation and you would have to search https://github.com/NVIDIA/cccl to find out API names though.

[nastya236]

Thanks, great question. My intention with raw PTX was to make sure I understand exactly what each instruction does. I felt that when wrapped in ccc up the details are hidden bit, so writing them in a raw ptx makes them explicit. I agree on docs point. That being said if you have a strong preference, I am happy to substitute it with cccl functions :)

[angeloskath]

@nastya236 is there any reason not to merge this? Will do a quick pass but I think it should be good to go, isn't it?

[nastya236]

The reason we have not merged it is that i was too slow to respond to Cheng's comment. I think it is mergable, thank you!

[angeloskath]

👍