Break the benchmarks to have separate executables

[muhammad-tanvir-1211]

This PR breaks the benchmarks into multiple micro-benchmarks (1 per operator e.g. GEMM, Flash Attention Prefill etc.). It also adds the option to build and run all the benchmarks by compiling only 1 executable file (ninja benchmarks) similar to the examples and tests.

[joeatodd]

LGTM, nice one 👍

I am trying to upskill in CMake a little, and saw an opportunity to make some improvements here. Since we are all in forks, the easiest way I could find to communicate my suggestion is to provide this branch.

The idea basically is to automate the dependency handling by defining benchmark 'suites'. Each subdir now defines its own suite (which may be part of a larger suite). I added a mandatory SUITE argument to cutlass_benchmark_add_executable, so it should be impossible to make a mistake when adding more content in future.

Perhaps its overkill, please feel free to ignore! Equally, I can open it as a separate PR later.

[aacostadiaz]

Thanks!!

Could you add a README file to the benchmark folder explaining how to run the benchmark?

[muhammad-tanvir-1211]

Thanks!!

Could you add a README file to the benchmark folder explaining how to run the benchmark?

Done in a0b2470