Support more CUDA API in CuPy JIT

Part of #4290

CuPy JIT needs more coverage of the functions and attributes supported in CUDA. Each function can be implemented in CuPy JIT by writing a class that inherits from BuiltinFunc here.

• Math modules: https://docs.nvidia.com/cuda/cuda-math-api/
  • math module functions
  • cupy ufunc
  • CuPy JIT specific interface
• Atomic Functions: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#atomic-functions
  • atomicAdd #5169
  • atomicSub #5387
  • atomicExch #5387
  • atomicMin #5387
  • atomicMax #5387
  • atomicInc #5387
  • atomicDec #5387
  • atomicCAS #5387
  • atomicAnd #5387
  • atomicOr #5387
  • atomicXor #5387
• Warp intrinsics: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#warp-shuffle-functions
  • Warp shuffle functions #5387
  • Warp match functions
  • Warp reduce functions

[asi1024]

@leofang I feel difficult to solve this problem from ThreadGroup class definition. I have opened a PR that combines declarations and initializations! (#6619)

[leofang]

xref: https://numba.readthedocs.io/en/stable/cuda-reference/kernel.html