Ensure compatibility with "-Wimplicit-fallthrough" when compiling with Clang

[wenxin0319]

This commit makes its files safe for use with "-Wimplicit-fallthrough".

[ccecka]

Good point, though we should prefer a solution that doesn't use exceptions or std::.

[github-actions]

This PR has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this PR if it is no longer required. Otherwise, please respond with a comment indicating any updates. This PR will be labeled inactive-90d if there is no activity in the next 60 days.

[r-barnes]

@ccecka - What solution would suggest?

Elsewhere in the project I see std::cerr followed by abort():

cutlass/tools/util/include/cutlass/util/device_rmsnorm.h

Lines 181 to 182 in b995f93

	std::cerr << "CUDA error: " << cudaGetErrorString(result) << std::endl;
	abort();

and std::cerr followed by std::exit:

cutlass/examples/41_fused_multi_head_attention/fused_multi_head_attention_backward.cu

Lines 89 to 90 in b995f93

	std::cerr << "FATAL: Read '" << read << "' but expected '" << expected << "'" << std::endl;
	std::exit(1);

and a number of uses of throw (link):

cutlass/tools/library/src/handle.cu

Line 62 in b995f93

throw std::runtime_error("cudaGetDevice() failed");

cutlass/tools/util/include/cutlass/util/device_memory.h

Line 69 in b995f93

throw cuda_exception("Failed to allocate memory", cuda_error);

As contributors @wenxin0319 and I would like to match the existing project style, and it seems to us that throw std::runtime_error is pretty reasonable given its existing use within the project.

If that's not the case, what should we do instead?

[thakkarV]

keeping live. will get back to you

[ccecka]

We would prefer static_assert and assert when possible, but since these are host-only I suppose it's fine.

Approving now.

[thakkarV]

@Junkai-Wu CC

[Junkai-Wu]

@hwu36 This PR is ready to merge.

[r-barnes]

I've tried to enable -Wimplicit-fallthrough in #2661, but I'm not sure what I'm doing.