Fix silent GPU transfer failures disabling CPU threading fallback

[bhuvan-somisetty]

Fixes #175

to_gpu() in osipy/common/backend/array_module.py caught any exception from the actual GPU transfer and fell back to NumPy with zero logging or warning. That's already not great, a real CUDA OOM or driver error just vanishes, but BaseFitter.fit_image() (osipy/common/fitting/base.py) makes it worse: use_gpu is decided once before the transfer and never re-checked afterward, so on a silent fallback:

• Chunk sizing still picks GPU-oriented sizing instead of the CPU chunk_size.
• use_threading = not use_gpu and ... stays False because use_gpu is still True, so the CPU multi-threaded fallback that would normally compensate is disabled too.

End result: a GPU transfer failure silently degrades fitting to serial, single-threaded CPU execution with nothing telling you why. osipy/common/backend/batch.py already treats an analogous GPU-OOM case as worth a warnings.warn(...), this just brings to_gpu() in line with that existing pattern.

Fix

• to_gpu() now warns (UserWarning) when the transfer fails, before falling back.
• fit_image() re-derives use_gpu from the actual array to_gpu() returns (hasattr(..., "__cuda_array_interface__")) instead of trusting the pre-transfer decision, so chunk sizing and the CPU threading fallback reflect what's actually running.

Testing

• Added TestToGpu::test_transfer_failure_warns_and_falls_back in tests/unit/common/backend/test_array_module.py: mocks a GPU transfer failure and confirms a UserWarning is raised and a usable NumPy array is still returned.
• Added tests/unit/common/fitting/test_base.py (new file, mirrors the osipy/common/fitting/ package): confirms that when to_gpu() falls back, fit_image() re-enables the CPU threaded path (previously it silently stayed disabled), plus a sanity check that a successful GPU transfer still skips CPU threading as expected.
• Verified both new tests actually fail against the pre-fix code (reverted the fix locally and re-ran, both failed with the expected assertion errors) before confirming they pass with the fix.
• Full suite: 778 passed, 40 skipped (GPU/data-dependent), 6 xfailed, 8 failed. Those 8 are the pre-existing PackageNotFoundError CLI failures already tracked/fixed separately in [BUG] osipy CLI crashes on every invocation when package metadata isn't discoverable, not just --version #173 / PR Fix CLI crashing on every invocation when package metadata is missing #174, unrelated to this change (this branch was cut from main before that fix landed).
• ruff check / ruff format --check clean on all changed/new files.
• mypy on both changed source files: no new errors introduced (verified by diffing mypy output against main, same 3 pre-existing findings in array_module.py shifted by one line from the added import; fitting/base.py is clean).

[bhuvan-somisetty]

cc @ltorres6 @MohamedNasser8 for review

[MohamedNasser8]

Hello @bhuvan-somisetty, thanks for all your work, I'll have a look soon.

[ltorres6]

Thanks for catching this silent fallback and your fix attempt.

I think what we actually want here is to throw an error instead of falling back if we
a) have a detected GPU
b) are not using force_cpu, and
c) transfer fails for any reason.

This will also require a change in batch.py since that deliberately falls back to numpy if we fill memory.

Can you make those changes?

[ltorres6]

Please use the logging package to log INFO/WARN/ERRORS.