REFACTOR: Shared image fetch helper for multimodal seed datasets

[romanlutz]

Summary

Follow-up to @jsong468's review comment on #1757: #1757 (comment) (out of scope for this PR).

Four multimodal seed-dataset loaders under pyrit/datasets/seed_datasets/remote/ duplicated the same download an image from a URL, cache it under seed-prompt-entries, return the local path logic. This PR extracts that logic into a single shared async helper.

What changed

Added pyrit/datasets/seed_datasets/remote/_image_cache.py with:

async def fetch_and_cache_image_async(
    *,
    filename: str,
    image_url: str | None = None,
    image_bytes: bytes | None = None,
    log_prefix: str = ""image-cache"",
    request_headers: Mapping[str, str] | None = None,
    request_timeout: float | None = None,
    follow_redirects: bool = False,
) -> str

• Caller passes filename so each loader keeps its existing on-disk cache filename pattern (no user-visible cache invalidation).
• image_bytes path supports loaders that have raw bytes in hand (e.g. PIL images from HuggingFace rows) without going through the network. This makes the helper a drop-in fit for the MSTS loader once FEAT: Add MSTS multimodal safety dataset loader #1757 lands.
• request_headers/request_timeout/follow_redirects cover VLSU's custom UA + 2 s timeout + redirect needs without leaking a generic **httpx_kwargs through the public API.
• Uses str(Path(results_path) / sub_dir / filename) for cross-OS-stable cache paths so the skip download if file exists fast-path fires correctly on Windows.
• Raises on failure to match the existing try / except Exception -> log + skip convention used at every call site.
• Preserves the existing RuntimeError message (Serializer memory is not properly configured) so existing pytest.raises(match=...) assertions keep passing.

Why a standalone module (rather than a base-class method or mixin)

• Only 4 of ~30 remote dataset loaders touch images - adding the method to _RemoteDatasetLoader would bloat the base.
• Standalone helper is easy to test in isolation.
• No existing _helpers.py / _common.py in remote/, so the new module fits the prevailing one-helper-per-file pattern.

Migrated loaders

Loader	Method	Filename pattern (unchanged)
_HarmBenchMultimodalDataset	_fetch_and_save_image_async	harmbench_{behavior_id}.png
_VisualLeakBenchDataset	_fetch_and_save_image_async	visual_leak_bench_{example_id}.png
_VLSUMultimodalDataset	_fetch_and_save_image_async	ml_vlsu_{group_id}.png (custom headers/timeout/redirects preserved)
_ComicJailbreakDataset	_fetch_template_async	comic_jailbreak_{template_name}.png (template-name validation kept in the wrapper)

Each loader's existing wrapper method survives as a 1-3 line delegating call, so every existing patch.object(loader, "_fetch_and_save_image_async", ...) test mock keeps working untouched.

_VLGuardDataset is not migrated - it downloads a zip from HuggingFace Hub instead of an individual image URL, so the pattern doesn't apply.

Tests

• New tests/unit/datasets/test_image_cache.py (8 tests) directly exercising the helper: cache hit skips network, URL fetch path writes response bytes, image_bytes path skips network entirely, ValueError on missing input, RuntimeError on misconfigured memory, exception propagation, custom headers/timeout/redirects forwarded to the HTTP client, and graceful fallback when the cache-existence check itself fails.
• Existing loader tests adjusted in three small ways:
  • Patch target switched from <loader_module>.data_serializer_factory to _image_cache.data_serializer_factory (the loaders no longer import the factory directly).
  • expected_path assertions switched to str(Path(...)) for Windows/POSIX portability.
  • No test logic changes.

All 439 unit tests under tests/unit/datasets pass.

Verification

uv run ruff format pyrit tests doc            # 1127 files left unchanged
uv run ruff check pyrit tests doc              # All checks passed!
uv run -m ty check pyrit                       # No new diagnostics in changed files
uv run pytest tests/unit/datasets -v           # 439 passed

[jsong468]

Thanks for addressing!