Keep interface validation active under python -O

[adenzhou1350]

Summary

• replace user-facing assert checks for unsupported begin/hint arguments
  with explicit validation that remains active under python -O
• reject a preallocated output_idx whose dtype disagrees with
  indices_type, also before output allocation or backend dispatch
• add a CPU-only regression test that replaces the CUDA extension with a
  minimal fake while exercising the real Python interface with PyTorch tensors

Testing

python tests/test_interface_validation.py
# 4 tests passed

python -O tests/test_interface_validation.py
# 4 tests passed

python -OO tests/test_interface_validation.py
# 4 tests passed

python -m py_compile deep_select/interface.py tests/test_interface_validation.py
python -m ruff check deep_select/interface.py tests/test_interface_validation.py
git diff --check origin/main...HEAD

The same regression test projected onto unmodified commit
0f03b68748b304863fdf0181a11458d04ae533a9 fails all three expected-rejection
cases under python -O: the removable assertions allow the fake backend to
receive unsupported arguments or a mismatched output dtype. Its positive
control still passes. The candidate rejects all three invalid calls before
backend dispatch and sends the valid call to the backend exactly once.

Scope

This is Python interface validation only. It does not change CUDA kernels,
supported argument combinations, numerical behavior, or performance.