Reduce Python CPU overhead and improve validation messages

[matthewdouglas]

Reduce Python dispatch overhead, particularly for the CUDA/ROCm backend, and improve DX with validation checks and error messages. Some overlap with #1408.

• CUDA/ROCm: Reduce ctypes overhead for C kernel calls (set argtypes/restype, other related improvement)
  • gemm_4bit fallback path: avoid redispatch via torch.ops, instead inline
• Convert torch._check to cheaper if/raise within custom op implementations
• Hoist many validation checks out of custom ops and into the public APIs where appropriate

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.