Enable CUTLASS grouped GEMM for pretraining wgrad on GB200 and H100 (resubmit) #18781

Job	Run time
build-macos (macos-13, static)	13m 51s
build-bazel (x86, linux.12xlarge, gcc)	10m 4s
build-bazel (x86, linux.12xlarge, clang)	9m 26s
build-macos (macos-13, shared)	23m 31s
build-windows (windows-2022, static)	12m 6s
build-windows (windows-2022, shared)	12m 33s
build-linux (x86, linux.12xlarge, shared, clang, 16.0.6)	19m 52s
build-linux (arm, linux.arm64.m7g.4xlarge, static, clang, 16.0.6)	33m 11s
build-linux (x86, linux.12xlarge, static, gcc, 9.5.0)	25m 59s
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 (arm, linux.arm64.m7g.4xlarge, shared, clang, 16.0.6)	32m 45s
build-linux (x86, linux.12xlarge, static, gcc, 14.1.0)	26m 18s
build-linux (x86, linux.12xlarge, shared, gcc, 14.1.0)	26m 31s
build-linux (arm, linux.arm64.m7g.4xlarge, shared, gcc, 14.1.0)	1h 24m 53s
build-linux (arm, linux.arm64.m7g.4xlarge, static, gcc, 14.1.0)	1h 20m 31s
	7h 37m 24s

Provide feedback