Cumm

Added
- add cuda-like kernel in apple silicon, only RTC compile is supported.
Changed
- BREAKING CHANGE: rename macro `TV_CUDA` to `TV_ENABLE_HARDWARE_ACC`
- BREAKING CHANGE: change managed tensor behavior, now you must specify cpu device when you use managed flag.

Fixed
- fix compile problem in cuda 12.x

Added
- add `run_in_process` support for inliner to debug some unrecoverable cuda errors such as invalid memory access (700) without restart whole process. this option will copy all tensor data to cpu, copy them to child process (spawn mode), run in child process, and copy back to cpu and main process. this will slow down the performance, but it's very useful for debugging.
- add macro `TV_ASSERT_WITH_PRINT` to perform print in assert.
- change inliner function name with user-provided name for debug.

Fixed
- fix a small bug in `mp_helper.h`

Added
- Add std flag to NVRTCInlinerBuilder
- add `get_nvrtc_kernel_attrs` to NVRTCInlinerBuilder
- add prompt for inliner, use `python -m cumm.inliner.cuda` or `python -m cumm.inliner.cpu` (clang must be installed)
- add rich message print support for nvrtc compile powered by awesome `rich` library. (don't support llvm)
Changed
- change nvrtc tuple impl to support std::tie
- change supported cuda version, remove cuda 10.2 and 11.6, add cuda 12.1 and 12.2
- remove python 3.7, add python 3.12.
Fixed
- fix a small bug when using c++17 in nvrtc

Added
- add simple perf tools