[ROCm] update rocm to 6.4.2

[SergeyFilippov]

Update ROCm dependencies to the latest stable version.

Highlights:

• This release added experimental support for "Memory Efficient Flash Attention".
  (noticeable performance improvements starting with the second run of workflow).

  Command line adjustment for activation:

  -e TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1

• Add official support for RDNA 4 (9060/9070) link

[YanWenKun]

Thank you! Merging :-)
I'll update the doc later, and test torch 2.7.1 + rocm 6.4.2.

[YanWenKun]

It seems that TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 is no longer necessary since PyTorch 2.4:

https://github.com/Comfy-Org/docs/blob/main/troubleshooting/overview.mdx

But in this doc I found another env var could tweak performance:

# Enable tunable operations (slow first run, but faster subsequent runs)
PYTORCH_TUNABLEOP_ENABLED=1 python main.py