move float8 blockwise kernels out of prototype

Summary:

These will be useful as a fallback path for a_1_128_w_128_128 (DeepSeek)
scaling support for float8 inference.  Bringing out of prototype folder.

Test Plan:

```
pytest test/kernel/test_blockwise_triton.py -s -x
python benchmarks/benchmark_blockwise_scaled_linear_triton.py
```

Reviewers:

Subscribers:

Tasks:

Tags:
ghstack-source-id: 0fb2407
ghstack-comment-id: 3460951786
Pull-Request: #3256

Author: vkuzo

benchmarks/benchmark_blockwise_scaled_linear_triton.py

@@ -13,7 +13,7 @@
     from triton.testing import do_bench
 
     from torchao.float8.float8_utils import compute_error
-    from torchao.prototype.blockwise_fp8_inference.blockwise_quantization import (
+    from torchao.kernel.blockwise_quantization import (
         blockwise_fp8_gemm,
         fp8_blockwise_act_quant,
         fp8_blockwise_weight_quant,

test/prototype/test_blockwise_triton.py renamed to test/kernel/test_blockwise_triton.py

@@ -11,7 +11,7 @@
 
 triton = pytest.importorskip("triton", reason="Triton required to run this test")
 
-from torchao.prototype.blockwise_fp8_inference.blockwise_quantization import (
+from torchao.kernel.blockwise_quantization import (
     blockwise_fp8_gemm,
     fp8_blockwise_act_quant,
     fp8_blockwise_weight_dequant,

torchao/prototype/blockwise_fp8_inference/blockwise_quantization.py renamed to torchao/kernel/blockwise_quantization.py

@@ -1,11 +1,12 @@
-from .blockwise_linear import BlockwiseQuantLinear
-from .blockwise_quantization import (
+from torchao.kernel.blockwise_quantization import (
     blockwise_fp8_gemm,
     fp8_blockwise_act_quant,
     fp8_blockwise_weight_dequant,
     fp8_blockwise_weight_quant,
 )
 
+from .blockwise_linear import BlockwiseQuantLinear
+
 __all__ = [
     "blockwise_fp8_gemm",
     "BlockwiseQuantLinear",

torchao/prototype/blockwise_fp8_inference/__init__.py

@@ -7,7 +7,7 @@
 import torch
 from torch import nn
 
-from torchao.prototype.blockwise_fp8_inference.blockwise_quantization import (
+from torchao.kernel.blockwise_quantization import (
     blockwise_fp8_gemm,
     fp8_blockwise_act_quant,
 )