…rib#1516)

The GPU decode path was wrapped in a too-broad ``try/except Exception:
pass`` that silently fell back to CPU on any error. Real GPU bugs (xarray-contrib#1508
was an AttributeError) lived undetected because the user-visible result
stayed correct.

Behaviour now:

- Default ``gpu='auto'`` keeps the CPU fallback but emits a
  ``RuntimeWarning`` that includes the original exception type and
  message. Stacklevel=2 so the warning points at the caller.
- New ``gpu='strict'`` re-raises the original exception instead of
  falling back. Useful for tests and CI that exercise the GPU fast
  path.
- Unknown ``gpu=`` values raise ``ValueError`` early.

Catch is still ``Exception`` rather than a narrower set: CUDA failures
arrive as many distinct types (``cuda.RuntimeError``,
``CudaAPIError``, ``CUDARuntimeError``, plain ``RuntimeError``, ...).
``BaseException`` is intentionally not caught so KeyboardInterrupt and
SystemExit propagate.

Test ``xrspatial/geotiff/tests/test_gpu_strict_fallback_1516.py``
monkeypatches ``gpu_decode_tiles_from_file`` to raise a synthetic
``RuntimeError`` and asserts the warn-and-fallback and strict-reraise
paths. Tests do not need a real GPU; a ``cupy`` shim is installed when
the real package is unavailable.

Closes xarray-contrib#1516

- Plumb gpu='strict'/'auto' through the second GPU decode stage too.
  Previously only gpu_decode_tiles_from_file failures honored the mode;
  the gpu_decode_tiles call after CPU-mmap tile extraction silently
  fell back regardless. Strict now raises and auto warns from either
  stage.
- Update the gpu kwarg docstring to describe the actual two-stage
  fallback (GDS GPU decode -> CPU-mmap + GPU decode -> full CPU + transfer).
- Strengthen _ensure_cupy_stub: also stub when cupy is installed but
  CUDA is not available, and restore the original module after each test.
- Add tests covering the second-stage strict re-raise and second-stage
  warning behaviour.