Fixes for LLVM 22, 23

[alexreinking]

This PR includes several fixes that were discovered when upgrading the buildbots to pin LLVM versions.

• NoInfsFPMath has been removed from LLVM main (23), so we now rely on FastMathFlags even in CodeGen_PTX_Dev
• LLVM's AArch64 AsmPrinter can stack overflow on macOS worker threads (which have tiny 512kb stacks), so we run JITModule::compile_module on a large stack now.
• On Windows, Ninja 1.13.0 with MSVC link.exe fails, apparently due to RSP line limit of 16,383 characters ninja-build/ninja#2616 breaks our build. So we prevent pip from installing 1.13.0 and hope that Kitware will update the fork they publish to PyPI to 1.13.1 or newer.