- Update building of openmp libraries for LLVM 20
- Add patches for standalone building and linking statically LLVM
- Use ninja for faster builds
- Add more printouts in setup
- Use build_wheel directly from setuptools since we require a recent
version

- Avoid multiple registration issues when using run_static_constructors
from the execution engine, which re-iterates every module

- Update libomptarget nvptx device RTL path
- Use the new pass manager
- Optimize toolchain code generation for the CUDA target

- Remove deprecetad nvvm.annotatios for CUDA kernel and use ptx_kernel
calling convention instead
- Replace startswith with starts_with
- Ddd default dummy argument to device kernel as expected by thee
runtime
- Set thread limit for cpu device(1) target regions
- Fix basic block move code to use eraseFromParent, insertInto
- Update tgt_register_lib and tgt_unregister_lib handling in ctor

- Use opaque pointer types
- Handle nodiscard APIs

- Include in the wheel only strictly necessary files via updating
MANIFEST.in
- Update setup.py to forward CC, CXX flags and avoid default cmake
system includes for manylinux building

- Expected by libomptarget to avoid a warning. Should not affect
correctness. Makes GPU execution more CPU-efficient by avoiding spinning
on GPU sync.

- Other packages (e.g., numba 0.62+, openblas) may require llvm-openmp and
bundling our own creates conflicst.

- Build libomp and libomptarget selectively to customize for different
packaging systems (conda or pypi) and platforms. We use the native
libomp and build libomptarget as needed to avoid conflicts.
- Rename libomp to openmp as it stores both libomp and libomptarget
built libraries
- Update python to refer to the renamed directory when finding
libraries.
- Raise an error if init expects to find a library

- Defined env vars to build bundled openmp libraries
- Use clang for libomptarget as required
- Add zstd, libffi dependencies to the conda environment