Performance update on the backward split kernel

[jtang10]

This PR improves the backward kernel performance, bringing it closer to the tutorial/06-fused-attention.py performance. As it shows above, based on the benchmark in flash_attn/flash_attn_triton_amdwe, we are on average 90% of the tutorial, and achieves 60% performance improvement from the previous PR #122.

The improvement comes from the following places:

1. tot Triton compiler, which greatly alleiviates the register spilling problem in the dkdv kernel.
2. Merged dkdv and dq kernel. There is nothing changed algorithm-wise, simply merging two kernels together to share common variables and reduce launch latency, like the tutorial does.
3. Turn on use_exp2 by default.