ci: add GitHub Actions test workflow across Python 3.8/3.10/3.12 (#28)

[ApoorvBrooklyn]

Add .github/workflows/test.yml: runs pytest (network tests disabled), ruff check, and mypy on every PR and push to main, matrixed across Python 3.8, 3.10, and 3.12. AGENTS.md previously documented "no CI" as a deliberate maintainer choice; this closes that gap with the lightest workflow that catches real cross-version regressions.

While validating the matrix locally (no 3.8 available, so verified 3.10/3.12 in throwaway venvs), found two real version-sensitive bugs the new CI would have caught immediately:

• unittest.mock.patch("...downloader.Downloader._DEFAULT_REGISTRY") resolves incorrectly on Python <=3.10: this package's init.py re-exports the legacy downloader() function under the same name as the downloader submodule, and mock's pre-3.11 string-target resolution walks that dotted path attribute-by-attribute instead of using importlib, landing on the function instead of the module. Python 3.11+ switched to pkgutil.resolve_name, which doesn't hit this. Fixed by importing Downloader directly and using patch.object() instead of a string target, in test_engine.py, test_download.py, and test_features.py (12 call sites).
• The two TestMultidownloaderDeprecated tests import crawler.py / multidownloader.py, which import selenium unconditionally; CI only installs the [dev] extra, not [google], so these failed on import. Gated both behind pytest.importorskip("selenium").

Also: add a CI badge to README.md, and update AGENTS.md's "no CI" note to describe the new workflow.

Closes #28

[KTS-o7]

Mixed feelings on this one. The test fix is real and worth merging; the CI workflow itself needs a policy decision first.

The patch.object / mock-target fix is correct and shippable on its own. The comment in test_engine.py documenting the pkgutil.resolve_name vs attribute-walk difference is exactly the kind of context future contributors need. Keep that comment verbatim if anything is split out.

However, the workflow itself has issues that need maintainer input before merge:

1. Reverses a documented policy. AGENTS.md (line 14) states: "No CI is configured (per the maintainer's request)." This PR closes Add a GitHub Actions workflow to run pytest + ruff + mypy on every PR (Python 3.8, 3.10, 3.12) #28, which is the issue requesting this workflow — i.e. it's a unilateral flip of the maintainer's stated preference. The maintainer should approve adding CI before the YAML lands. A PR that contradicts the contributing guide is something to hash out in an issue first.

2. Python 3.8 is EOL (Oct 2024). Running it in CI to "match the project's documented floor" is wasted minutes. Either bump the documented floor to 3.9+ or drop 3.8 from the matrix.

3. Matrix gaps. The 3.8 → 3.10 → 3.12 jump skips 3.9 and 3.11. If the goal is "catches real cross-version regressions," 3.11 is the LTS-shaped gap most likely to catch the very kind of mock-import bug this PR fixed. Worth including.

4. black excluded arbitrarily. AGENTS.md says pre-commit runs black/ruff/mypy. The workflow runs ruff + mypy only. Either include all three (matches the project's stated lint stack) or document why one is dropped.

5. Selenium gate perpetuates deprecated code. pytest.importorskip("selenium") in test_engine.py keeps the deprecated crawler.py / multidownloader.py tests alive in CI. Those modules are slated for removal in v4.0.0 (per AGENTS.md), and the deprecation is already encoded in their docstrings. The cleaner move is to delete those two tests rather than wire up selenium coverage for modules that shouldn't be tested in CI.

Suggested split:

• PR #54a (mergeable now): Tests only — the patch.object conversion across test_engine.py, test_download.py, test_features.py. Tiny, targeted, no policy question.
• PR #54b (after maintainer approves Add a GitHub Actions workflow to run pytest + ruff + mypy on every PR (Python 3.8, 3.10, 3.12) #28): Workflow + AGENTS.md + README badge. Open an issue first if needed to confirm the maintainer actually wants CI now.

The CI work is well-intentioned and the matrix comment about cross-version regressions is fair — it's just that the timing/policy question blocks the merge.