Add support for AMD GPU (ROCm Platform)

[yaomingamd]

PyPI name of openAI Triton for ROCm Platform is pytorch-triton-rocm, so that modify setup.py to install correct Triton package for ROCm platform. Also modify README.md to add instruction to install on ROCm Platform. Tested on ROCm Platform with AMD GPUs.

[KnechtNoobrecht]

What OS is this for?

[yaomingamd]

OS is Linux

[threeal]

Instead of using environment variable, i suggest to use extras_require as could be see in here. Another option is to automatically detect if it is using ROCm platform.

[yaomingamd]

Based on the suggestion, remove environmental variable and add function to detect ROCm Platform automatically.

[x86Gr]

Will this work with generic AMD gpus, ie newer integrated gpus?

[glangford]

Will this work with generic AMD gpus, ie newer integrated gpus?

@x86Gr This is the list of supported GPUs

https://rocm.docs.amd.com/en/latest/release/gpu_os_support.html

[yaomingamd]

@x86Gr @glangford This link should work: https://rocm.docs.amd.com/en/latest/release/gpu_os_support.html#linux-supported-gpus

[ic4-y]

Any particular reason the PR only selected a subset of supported AMD GPUs?

gfx1030 or gfx1100 appear to be missing

[yaomingamd]

@vadimkantorov @Reviewer of this PR. Is there any update about review to merge this PR? Is there anything I can help to speed up the process? Thanks!