Spconv

Fixed
- Fix build problem
- Fix nvrtc problem

Added
- Add Ampere support. faster fp16, faster tf32 and greatly faster int8 kernels in Ampere GPUs.
- Add pure c++ code generation (libspconv.so) for deploy (or train in another deeplearning framework)
- Add NVRTC support for all gemm kernels. if your GPU architecture isn't compiled in prebuilt, spconv will use slightly slower (10-20us overhead for every kernel launch) NVRTC kernels.

Fixed
- Fix launch fail in maxpool if too much voxels

Changed
- all weight layout will be KRSC, don't support old spconv 1.x weights anymore.
- previous gemm ops in ops.py now move to c++ by default (controlled by spconv.constants.SPCONV_CPP_GEMM)

Removed
- drop python 3.6 support.
- pascal and kepler architecture is removed in CUDA 12 prebuilt.

Added
- add full nvrtc support
- add support for large spatial shape and batch size. if detect large shape, we use int64 instead of int32 when hashing.

Added
- add sm_37
- add fp16 kernels witl fp32 accumulator (run slower, but can avoid nan if channel size is too large)
- add SPCONV_BWD_SPLITK env to control splitk candidates.

Added
- Add fp16 conv simt kernels for mixed-training in pascal or older GPUS. WARNING: not optimized for TESLA P100 which has 2x throughput in half.

Fixed
- Fix wrong arch assert in all kernels for old GPUs to make spconv work in sm_50 GPUs