fix(generators): raise clear error on HF InferenceAPI 404 response

[pujitha24]

Tell us what this change does. If you're fixing a bug, please mention
the github issue number.

This improves the error message raised by huggingface.InferenceAPI when the
Hugging Face Inference API returns an HTTP 404 for a model. Previously this
surfaced as a generic TypeError: Unsure how to parse 🤗 API response type: b'Not Found', please open an issue..., which told users to file a bug report
for what is actually an HTTP-level "model not found" response. This is
exactly what happened in issue #1297: a maintainer identified that the model
in question is not exposed via the legacy default Inference API provider (HF
introduced a separate "Inference Providers" architecture in Jan 2025), but
the generator gave no indication of that — it just told the user to file a
bug.

Fully supporting HF's new Inference Providers routing (provider selection,
new endpoints/auth) would be a larger redesign of this generator and is out
of scope here. This change instead makes the immediate failure mode clear and
actionable: a dedicated ModelNotFoundError is now raised on HTTP 404,
naming the model and pointing at HF's Inference Providers docs, instead of
the misleading generic parse-error message.

Please ensure you are submitting from a unique branch in your repository to main upstream.

Verification

List the steps needed to make sure this thing works

• Supporting configuration such as generator configuration file — not needed; change is exercised via mocked HTTP responses in unit tests, no live generator config required.
• garak -t <target_type> -n <model_name> — not run against a live Hugging Face endpoint (would require a real HF_INFERENCE_TOKEN and a model that 404s under the legacy API); the new code path is covered by a targeted unit test instead (see below).
• Run the tests and ensure they pass python -m pytest tests/: ran python -m pytest tests/generators/test_huggingface.py -v (all 8 tests pass, including the new test_inference_model_not_found). Confirmed the new test fails without the fix (AttributeError: module 'garak.generators.huggingface' has no attribute 'ModelNotFoundError') and passes with it, i.e. it reproduces the reported failure mode and proves the fix. Also ran black --check and a targeted pylint (undefined-variable/unused-variable/unused-import) pass on the two changed files — both clean.
• Verify the thing does what it should — new test test_inference_model_not_found mocks a 404 response with body b"Not Found" (matching the exact response shape from the issue's traceback) and asserts ModelNotFoundError is raised with a message identifying the model.
• Verify the thing does not do what it should not — confirmed ModelNotFoundError is not in the @backoff.on_exception retry list, so a 404 is not needlessly retried (a missing/unavailable model is not a transient condition, unlike the existing 503/rate-limit handling it sits alongside).
• Document the thing and how it works — the raised error message itself documents the likely cause and links to https://huggingface.co/docs/inference-providers for further reading; no separate doc page seemed warranted for a single, self-explanatory error message.

If you are opening a PR for a new plugin that targets a specific piece of hardware or requires a complex or hard-to-find testing environment, we recommend that you send us as much detail as possible.

Specific Hardware Examples:

• GPU related
  • Specific support required cuda / mps ( Please not cuda via ROCm if related )
  • Minium GPU Memory

Complex Software Examples:

• Expensive proprietary software
• Software with an extensive installation process
• Software without an English language UI

Report: #1297

---

AI assistance: this change was drafted with Claude Code.

Fixes #1297