IFU to v2.0.4

[jayz0123]

• [x] renamed "unpadded" -> "varlen" with mha_varlen_fwd & mha_varlen_bwd APIs.
• [x] changed mha_fwd & mha_bwd for input with the same sequence lengths in the same batch.
• [x] setup.py will install for either CUDA/ROCm system.
• [x] rename test_flash_attn -> test_flash_attn_rocm for ROCm unit test.
• [x] benchmark testing
• [x] Sync to PR #15
• [x] add unit tests for mha_fwd&mha_bwd
• [x] MQA/GQA
• [ ] unit test all pass

Current Unit Test Result: (PyTorch 2.0.0; ROCm 5.6) 3968 passed, 63 skipped

Current Performance on MI250: (docker pull rocm/pytorch:rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1)	fwd tflops	bwd tflops	total
fp16	52.16	39.93	42.49
bf16	52.36	30.25	34.21

[jayz0123]

A new environment variable "FLASH_ATTENTION_INTERNAL_ENABLE_TIME_KERNEL" can switch the output of kernel running time

[jayz0123]

[BUGs] Previously in older version of FA, we create tensors z and softmax_lse matrix of max sequence lengths with no padding for grouped gemm. But the strides for each batch for the tensors are different. This behaviour will cause wrong result from CK. Fixing it.

[fsx950223]

Please remove *_hip.hpp