Improve API for f4f4bf16 (#5163)

Summary:
Pull Request resolved: #5163

X-link: https://github.com/facebookresearch/FBGEMM/pull/2162

We add some improvements for FP4 gemm.
- Remove the need to pass `use_mx`, we can infer this based on `global_scale`
  - As a follow up we should improve the assertions on the proper FP4 dtypes, similar to what we have with [FP4 group gemm](https://www.internalfb.com/code/fbsource/[addad803d330]/fbcode/deeplearning/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/cutlass_extensions/f4f4bf16_grouped.cu?lines=388-409).
- Add optional `output` to API, which is in-line with other torch APIs.
- Move function declaration to `torch_ops.h` which will remove the need for the forward declaration in Blas.cpp
- Small code cleans ups

Misc
- Later we should likely clean-up & re-evaluate the heuristic for the kernel, right now its almost identical (and duplicated) for MX and NV FP4, and we are likely instantiating more instances than needed.

Reviewed By: slayton58

Differential Revision: D87655845

fbshipit-source-id: d3ddd1f7efe7683fb8615ab2f6febe438ce6b380

Author: cthi

fbgemm_gpu/experimental/gemm/triton_gemm/fp4_quantize.py

@@ -289,7 +289,7 @@ def triton_quantize_mx4_unpack(
         stochastic_casting (bool): Whether to use stochastic casting.
 
     Returns:
-        torch.Tensor: [M / 2] mx4 scaled tensor packed into in8
+        torch.Tensor: [M / 2] mx4 scaled tensor packed into uint8
         torch.Tensor: [M / group_size] mx4 shared exponents into int8
 
         eg.

fbgemm_gpu/experimental/gen_ai/bench/quantize_ops.py

@@ -2385,7 +2385,7 @@ def quantize(self, x, w):
 
     def compute(self, xq, wq, x_scale, w_scale, global_scale):
         return torch.ops.fbgemm.f4f4bf16(
-            xq, wq, x_scale, w_scale, global_scale=global_scale, use_mx=False
+            xq, wq, x_scale, w_scale, global_scale=global_scale
         )
 
     def quantize_and_compute(self, x, w):
@@ -2471,7 +2471,7 @@ def quantize(self, x, w):
 
     def compute(self, xq, wq, x_scale, w_scale, global_scale):
         return torch.ops.fbgemm.f4f4bf16(
-            xq, wq, x_scale, w_scale, global_scale=global_scale, use_mx=False
+            xq, wq, x_scale, w_scale, global_scale=global_scale
         )
 
     def quantize_and_compute(self, x, w):