[WIP] MUSA: enable fastfp16, correct warp reduce impl and perf tuning

[BodhiHu]

Make sure to read the contributing guidelines before submitting a PR

[WIP] This PR will:

1. enable fastfp16;
2. correct warp reduce impl;
3. perf tuning ...

[JohnLoveJoy]

The auto-label system is bugged, accidentally tagging everything with "Nvidia GPU."

[zhouwg]

how to requested a review from a specified expert in a specified PR? thanks so much!

[JohannesGaessler]

I can review the code in terms of the effects for the CUDA backend but I don't know anything about MUSA and do not have any MT hardware so I would need someone else to review this PR in terms of whether the changes are correct for MUSA.

[IMbackK]

Same here but for effects on the hip backed.

One thing code style wise i can say is that this pr currently contains alot of stuff that dose not belong like commented out logging printfs an the like.

Otherwise it looks reasonable but i have no way of verifying correctness for MUSA

[BodhiHu]

Hi, thanks for the comments here.
It's still a draft PR WIP, and not ready for review yet.

Sorry if it contains lots of debug prints now, which will be cleared when the perf tuning is done ;D