Improve CUTLASS GMM for llama4x pretraining fprop #18608

Job	Run time
build-macos (macos-13, shared)	33m 13s
build-bazel (x86, linux.12xlarge, gcc)	10m 8s
build-bazel (x86, linux.12xlarge, clang)	9m 23s
build-linux (x86, linux.12xlarge, static, clang, 16.0.6)	20m 9s
build-linux (x86, linux.12xlarge, shared, gcc, 9.5.0)	25m 44s
build-linux (x86, linux.12xlarge, static, gcc, 9.5.0)	25m 48s
build-linux (x86, linux.12xlarge, static, gcc, 14.1.0)	26m 18s
build-linux (arm, linux.arm64.m7g.4xlarge, static, gcc, 14.1.0)	1h 21m 1s
build-linux (x86, linux.12xlarge, shared, gcc, 14.1.0)	26m 30s
build-linux (x86, linux.12xlarge, shared, clang, 16.0.6)	19m 51s
build-linux (arm, linux.arm64.m7g.4xlarge, shared, clang, 16.0.6)	33m 6s
build-windows (windows-2022, static)	11m 56s
build-linux (arm, linux.arm64.m7g.4xlarge, shared, gcc, 14.1.0)	1h 22m 30s
build-linux (arm, linux.arm64.m7g.4xlarge, static, clang, 16.0.6)	32m 53s
build-macos (macos-13, static)	19m 31s
build-windows (windows-2022, shared)	12m 29s
	7h 50m 30s

Provide feedback