e2e_eval: dynamic tasks, priority/group cascade, --priority list

[xieofxie]

Summary

build_registry.py

• Fetch pipeline tasks dynamically from HuggingFace /api/tasks (47 tasks) instead of the 21-entry hand-written NLP_TASKS + CV_TASKS lists.
• 3-level cascade for priority and group when classifying entries:
  1. Curated — models_curated.json overrides win verbatim (group, priority both honored — previously priority was forced to P0).
  2. Existing — if (hf_id, task) was in the previous models_all.json, inherit that row's priority/group. Lets hand-edits to models_all.json survive rebuilds.
  3. New-rule defaults — microsoft/-prefixed → (microsoft, P1); everything else → (Top200, P2). (P2 is a new tier; optimum_supported no longer drives priority.)
• Phase 1.5 preservation: existing (hf_id, task) rows that aren't re-picked by Phase 1 are kept verbatim except downloads + optimum_supported are refreshed. Lets the registry stay stable across runs and supports tasks HF's API doesn't expose (e.g. masked-lm).
• Per-task order field: every entry now carries order: 1..N ranked by downloads descending so consumers don't need to re-sort.

run_eval.py

• --priority now accepts multiple values — e.g. --priority P0 P1 runs the union. Backward-compatible with single value.
• filter_registry() widened: priority: str | Sequence[str] | None.

Curated list

• Add 4 ISV models to models_curated.json.

Regenerated models_all.json (439 entries, sorted by (hf_id, task))

• Priority: P0: 22 | P2: 202 | P3: 215
• Groups: Top200: 398 | Foundry Toolkit: 16 | microsoft: 15 | Benchmark: 6 | ISV: 4
• Optimum-supported: 311/439