Fix long MSVC linker commands on Windows

[jammm]

Similar fix as pypa/distutils#406

Bringing this fix here as new distutils updates may take time to percolate across the community. This PR will immediately unblock users who would like to build flash-attention on RDNA3/4 on windows (#2054)

[jammm]

For anyone looking to build flash-attention on Windows on your RDNA3/3.5/4 GPU while this PR is awaiting to be merged, follow the commands below:

# Clone the PR branch:
git clone --recursive https://github.com/jammm/flash-attention -b jam/fix_long_linker_paths_windows

cd flash-attention

# Activate Visual Studio environment
cmd /c '"C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvars64.bat" >nul 2>&1 && set' | ForEach-Object { if ($_ -match '^([^=]+)=(.*)$') { [System.Environment]::SetEnvironmentVariable($matches[1], $matches[2], 'Process') } }

# Activate the virtual environment
.\venv\Scripts\Activate.ps1

# Set ROCm paths using rocm-sdk
$ROCM_ROOT = (rocm-sdk path --root).Trim()
$ROCM_BIN = (rocm-sdk path --bin).Trim()
$env:ROCM_HOME = $ROCM_ROOT
$env:PATH = "$ROCM_ROOT\lib\llvm\bin;$ROCM_BIN;$env:PATH"

# Set compiler and build settings
$env:CC = "clang-cl"
$env:CXX = "clang-cl"
$env:DISTUTILS_USE_SDK = "1"

# install flash-attention
pip install --no-build-isolation -v .

[jammm]

@rocking5566 @tridao PTAL. This will fix a long-standing issue of compiling flash-attention with CK backend on Windows. Once distutils PR gets merged and percolates, we can revert this change. But that will take a long time, so this solution is best to be merged in the short-medium term.