Break up cutlass_extensions.cu, pt1 (#2944)

Summary:
X-link: facebookresearch/FBGEMM#47

Pull Request resolved: #2944

Break up cutlass_extensions.cu

Currently,  `//deeplearning/fbgemm/fbgemm_gpu/experimental/gen_ai:quantize_ops_cuda#compile-cutlass_extensions.cu` is the second most expensive target in fbcode, at 1.77 EYS (see https://www.internalfb.com/intern/unidash/dashboard/ads_serving_platform/build_speed/).

As such, this is the first of N diffs to break up the source file into smaller chunks to parallelize the compilation and reduce the build times.

Before this diff - https://www.internalfb.com/buck2/4569ea23-35d0-4086-b7e1-ba30f1e665f8: (24:12.8s)
After this diff - https://www.internalfb.com/buck2/fffc0745-446d-4b22-bff9-3bff0c74eb58 (22:58.5s)

Reviewed By: jianyuh, spcyppt

Differential Revision: D60874248

fbshipit-source-id: 345ef2e579b391049aee8adf210d24b6e3190775

fbgemm_gpu/experimental/gen_ai/CMakeLists.txt

@@ -15,6 +15,7 @@ set(fbgemm_sources_include_directories
   ${CMAKE_CURRENT_SOURCE_DIR}/../..
   ${CMAKE_CURRENT_SOURCE_DIR}/../../include
   ${CMAKE_CURRENT_SOURCE_DIR}/../../../include
+  ${CMAKE_CURRENT_SOURCE_DIR}/src/quantize
   # PyTorch
   ${TORCH_INCLUDE_DIRS}
   # Third-party
@@ -31,6 +32,8 @@ set(attention_ops_sources
 
 set(quantize_ops_sources
     src/quantize/cutlass_extensions.cu
+    src/quantize/cutlass_extensions/i8i8bf16.cu
+    src/quantize/cutlass_extensions/i8i8bf16_dynamic.cu
     src/quantize/quantize.cu
     src/quantize/quantize.cpp)