[ROCm] Track latest composable_kernel

[a4lg]

The latest version of composable_kernel supports more versatile architectures and no longer assumes wavefront size of 64 (where many compiler errors come from).

This commit updates the composable_kernel submodule along with necessary interface changes.

This is a part of my attempt to support Flash Attention for Strix Halo (gfx1151) and it seems... CK portions used by Flash Attention already supports this architecture (not entire CK, though). I believe that my interface changes are fine (only adds/removes defaults).

p.s.
If someone has an AMD hardware already supported by Flash Attention (i.e. AMD Instinct), can you check the test results before and after this PR? In my Strix Halo environment, about half of the tests fail due to high arithmetic errors and I'd like to see whether this behavior is Strix Halo-specific.
If that's not Strix Halo-specific, I'll submit a follow-up PR to allow versatile AMD GPU architectures (to setup.py; possibly RDNA 2 or later?).

[tridao]

@rocking5566 does this interface change affect any existing case?

[rocking5566]

@a4lg could you also change c++ version in setup.py from c++17 to c++20?
Because lastest CK use c++20 by default

[rocking5566]

Actually, we are doing the similiar thing.
just finish the testing and about to send the PR from this branch.
https://github.com/ROCm/flash-attention/tree/ck_improve_v0.1.8

[a4lg]

@rocking5566
As long as CK is correctly updated, I don't stick to my changes.
BTW, my updated branch with suggested changes will be ready tomorrow (because I'm on a trip).

[rocking5566]

@rocking5566 As long as CK is correctly updated, I don't stick to my changes. BTW, my updated branch with suggested changes will be ready tomorrow (because I'm on a trip).

I submit the PR and change the c++ version.
#2052

We also test correctness and performance in this version (commit id) of CK in both MI300 and MI350.

[a4lg]

Closing in favor of #2052.