Update on "bring back torch.autograd.Function for float8 matmul"

Summary:

This is a redo of
#316

With upcoming support of scaling granularities other than tensorwise,
we need a good way to control which gemm kernel to call and how to scale
the input tensors in fwd and bwd. A `torch.autograd.Function` override
is the cleanest way to do that, and in 2024 this now works with
`torch.compile`.

Test Plan:

```
./test/test_everything.sh
```

Reviewers:

Subscribers:

Tasks:

Tags:

[ghstack-poisoned]

float8_experimental/config.py

@@ -84,7 +84,7 @@ class Float8LinearConfig:
     #
     gemm_config_output: Float8GemmConfig = Float8GemmConfig(use_fast_accum=True)
     gemm_config_grad_input: Float8GemmConfig = Float8GemmConfig()
-    gemm_config_grad_weight: Float8GemmConfig = Float8GemmConfig(use_fast_accum=True)
+    gemm_config_grad_weight: Float8GemmConfig = Float8GemmConfig()
 
     #
     # Per-linear configuration