test: fix stale/wrong test assertions and CI trigger gap

[musicalplatypus]

Summary

Two independent test/CI hygiene fixes, confirmed unrelated to any production code change (verified via git stash — all of the below fail identically on unmodified main).

Stale/wrong test assertions

• test_config_validation.py: KNOWN_TASK_TYPES and the model-registry check only covered timeseries + vision, so every audio modelzoo example (task_type='audio_classification', e.g. DSCNN_NPU) looked invalid. Includes the audio module's task types and routes audio configs to its own model registry.
• test_constants.py: get_default_data_dir_for_task() now raises ValueError on unknown categories, but the test still asserted the old silent DATA_DIR_CLASSES fallback. Updated to match current behavior.
• test_cross_device.py: TestQuantizationFlags called get_skip_normalize_and_output_int(..., partial_quantization=...), but the real parameter is auto_quantization — a wrong kwarg name, not a real API. Renamed calls; the one test that was actually about auto_quantization semantics is kept meaningful.
• test_dataset_utils.py: test_split_factor_too_large matched a stale error-message regex ("less than 1") against the current message ("must be in the range (0.0, 1.0)"); behavior was already correct, only the assertion was stale.

One gap found and left open, not guessed at: F28E12 is listed in TARGET_DEVICES but has no entry in _DEVICE_PROFILES, so test_cross_device.py::TestCompilationProfileCorrectness still fails for it (5 tests). This needs real hardware values (cross_compiler path, target, target_c_mcpu, has_hard_npu) that aren't safe to fabricate — flagging it here rather than papering over it with placeholder values.

CI trigger gap

tinyml-modelmaker depends on tinyml-modelzoo (pip install -e tinyml-modelzoo in the Install dependencies step), but dependency-only changes to tinyml-modelzoo wouldn't trigger the test suite at all. Adds the missing path trigger.

Note: this touches .github/workflows/test-modelmaker.yml, which two of my other open PRs (pr/dataloader-pinning-fix, pr/mps-eval-fixes) also touch — each adds one path-trigger line at the same position. If merged after either, GitHub will show a trivial 3-line conflict; the correct resolution is just keeping all the added lines (a union), not choosing one over the other.

🤖 Generated with Claude Code

[musicalplatypus]

CI failure is the separately-documented F28E12 gap, already called out in this PR's description — not introduced by this change.

test_cross_device.py::TestCompilationProfileCorrectness fails on F28E12: it's listed in TARGET_DEVICES but has no entry in _DEVICE_PROFILES. Left open deliberately since it needs real hardware values (cross_compiler path, target, target_c_mcpu, has_hard_npu) that aren't safe to guess at.

Worth noting this PR's own fix does work as intended: its run shows zero google_speech_command/audio-classification failures — https://github.com/TexasInstruments/tinyml-tensorlab/actions/runs/30575996012 — unlike every other currently-open PR (#18, #20, #21, #22, #23), which all still show those same 2 pre-existing failures (confirmed present on upstream/main itself, unmodified: https://github.com/TexasInstruments/tinyml-tensorlab/actions/runs/30376872256) until this one merges.