better install: streamline GPU detection and pip configuration

[Apophis3158]

I tried install.bat and found some issues:

Necessary

In TheRock/pytorch/README it says

With passing builds we upload torch, torchvision, torchaudio, triton, and apex wheels to subfolders of the "v2-staging" directory in the nightly release s3 bucket with a public URL at https://rocm.nightlies.amd.com/v2-staging/

Only with passing Torch tests we promote passed wheels to the "v2" directory in the nightly release s3 bucket with a public URL at https://rocm.nightlies.amd.com/v2/

So /v2 is tests passed stabled releases and /v2-staging should not be used.

Run pip install rocm[devel,libraries] first will always install the latest build like rocm-sdk 7.13.0a20260414, but then run pip install torch might only get torch 7.13.0a20260409, this will uninstall the rocm-sdk 7.13.0a20260414 you just installed and install rocm-sdk 7.13.0a20260409 instead.

• Fix: Run pip install torch torchaudio torchvision first, it will auto collect needed rocm-sdk core and libraries, then run pip install rocm[devel,libraries] will add the matching devel

When installation failed, the script prints Check the error messages above for details. but there was no error message.

• Fix: Remove several unnecessary >nul 2>&1

echo ! won't print because of setlocal enabledelayedexpansion set.

• Fix: use echo ^^! instead.

User might have set PIP_REQUIRE_VIRTUALENV=true for safety, and block script running,

• Fix: set "PIP_REQUIRE_VIRTUALENV=false" in script.

QOL

• bitsandbytes supports Windows ROCm CI now [ROCm] Windows workflow for creating wheels with ROCm 7.2.1 support bitsandbytes-foundation/bitsandbytes#1915 so we could use the nightly build via https://github.com/bitsandbytes-foundation/bitsandbytes/releases/tag/continuous-release_main.

• detect_gpu.py already got a lot of GPU information so we can just add an arch_index map to make better use of those.

• Use PIP_NO_WARN_SCRIPT_LOCATION and PIP_QUIET in script for quick environment installation debug.

[Apophis3158]

Run pip install torch first also fix another issue: If the user has set up a PIP mirror index, pip install rocm[devel,libraries] might only get a initial version rocm-0.1.0.tar.gz from mirror, because rocm-7.13.0a20260415.tar.gz requires --pre.