PyBind11 interface: How should function signature and CUDA error handling be handled for registered kernels?

I would like to understand the recommended approach when a Python call reaches these C++/CUDA interfaces.

Specifically:

How are the C++ function signatures validated by PyBind11?
What happens if Python passes tensors with an incorrect dtype, shape, device, or layout?
Where are the input validations performed — PyBind11 interface, C++ interface, or CUDA kernel?
How are CUDA errors propagated back to Python?
If a CUDA kernel encounters an illegal memory access, how is that error detected and reported through this interface?
Are there any recommended PyBind11 practices for preventing invalid tensor arguments from reaching the CUDA kernels?
Why are the interfaces exposed directly with m.def() rather than using explicit argument names, type annotations, or validation wrappers?

Review the PyBind11 registration points using m.def() and the associated C++/CUDA interfaces. Trace where function signatures, tensor validation, and CUDA errors are handled, then document the current behavior and recommended practices for invalid arguments and kernel failures.

Written by the indexing model from the issue text.