refactor(tests): extract shared HFLoadErrorTestsMixin (−247 LOC, +20 tests)

[stwilt]

Follow-up to #21. The four backend test files each hand-rolled a near-identical TestXBackendLoadErrors class (~180 lines for image_classification/object_detection, ~45 lines for grounding_dino/owlvit with only 3 of the 13 branches covered). Now that hf_load_error_handler is shared, the tests covering its branches should be shared too.

• Add HFLoadErrorTestsMixin to tests/_hf_test_utils.py holding all 13 load-error test methods, parameterized by class attrs backend_cls, processor_patch_target, model_patch_target, task_phrase.
• Replace four bespoke test classes with 5-line subclasses. As a side effect, grounding_dino and owlvit gain the 10 branches they were previously missing (ImportError allowlist, GatedRepo, HfHubHTTPError 401/non-401, Entry/LocalEntry, fallback, cause chain, model-side patch).
• Strengthen two assertions that were flagged in review of fix(backends): surface real cause on HF model load failures #21: (a) test_hf_hub_http_error_includes_repr now passes status_code=503 so label matches behavior, and asserts "503 Service Unavailable" appears in the message — catches any regression that drops repr(e). (b) test_local_entry_not_found_error_message / test_entry_not_found_error_message assert the cache-specific substring "not found in cache" rather than just "cache", so a future reorder that routes LocalEntryNotFoundError through the ValueError branch (it inherits from both) would flip the tests red instead of silently passing.

Net: +205 / −452 = −247 lines across 5 files. Load-error tests grow from 32 to 52 (4 backends × 13 branches). Full pytest run: 113 passed.

[lucasmundim]

Clean, well-executed DRY refactoring

The mixin pattern is correct (MRO ordering, no setUp/__init__, all parameterization via class attrs), the coverage gain is real (grounding-dino and owlvit jump from 3 to 13 branches), and the two assertion improvements are welcome:

• test_hf_hub_http_error_includes_repr now passes status_code=503 instead of defaulting to 404 while the message said "503".
• Entry/LocalEntry tests assert "not found in cache" instead of just "cache", catching regressions where an entry might route through the wrong handler branch.

Import cleanup is correct across all four files. Net -247 lines, test count 32 to 52. A rare case where removing code increases coverage.

Three minor non-blocking observations inline.

[lucasmundim]

Non-blocking: is a mutable class-level dict shared across all subclasses. No test mutates it today, so this is safe — but if a future test does it would leak across all subclasses. A defensive option:

Not urgent, just a prophylactic note.

[lucasmundim]

Non-blocking, just noting the shift: The old grounding-dino and owlvit each had an inline . The mixin drops that here and instead covers cause-chaining via the dedicated (ValueError path) and (model-side path). Since uses uniformly on all branches, the coverage is equivalent.

[lucasmundim]

Non-blocking: No CI checks have run on this branch yet ( reports nothing). Worth verifying the full 113-pass run before merging.