New GEMM kernels for weight-only quantization

[lzhangzz]

f16*u4g128 vs cublasGemmEx f16*f16, both using HMMA + f32 accumulator, on 32 weight matrices from 8 models range from 7B to 72B

• sm90 features are not used yet
• sm70 tensor core and fp16 have a shared pipeline

[lvhan028]

The main branch has OpenGVLab/InternVL2-40B-inner-4bits issue. This PR has nothing to do with it. @zhulinJulia24

[zhyncs]

Verified LMDeploy fp16 and AWQ on H100 with @lzhangzz , both are blazing fast.

[zhyncs]

TODO:

• Set the default max-batch-size to 256 when VRAM > 40G
• Resolve the runtime issue on SM90 by disabling certain code sections
• When using Quant, consider supporting the H100 feature (which can be implemented in other PRs)

[zhyncs]

Overall LGTM and just need to solve some minor issues with the H100 mentioned above.