fix(tasks): utils improvements across 7 benchmarks

[Luodian]

Summary

Generic, self-contained improvements to seven evaluation task utilities. No new dependencies; all data still loads from the existing public Hugging Face datasets.

Changes

• egoschema: Fix the HF_HOME cache path (~/.cache/huggingface, not .../hub). Replace the unbounded submission retry loop with a bounded, backed-off POST to EgoSchema's public validation server that saves the submission first and fails loudly on persistent failure instead of spinning forever or recording a bogus 0. Return the parsed accuracy from the MC / MC_PPL aggregators. Add deterministic generation_kwargs. README now spells out which tasks score offline (the N=500 subset) vs. which are submission-only (the full set) — and the offline subset tasks (egoschema_subset, egoschema_subset_mcppl) no longer declare the server submission metric at all: they have real answers and score locally, so an offline/no-egress node can't crash on the (now loud) server POST. The full-set tasks keep server submission.
• erqa: Use the shared extract_mcq_answer helper, add a doc_to_messages builder that interleaves images before the question text, and remove the dead local extractor / unused template load. Adds unit tests.
• vsibench: Robust numeric-answer extraction for NA-type questions — handles bare digits, digits embedded in a sentence, and English number words (e.g. twenty-one, three hundred and twenty-one), with the first number token winning.
• vsisuper: Harden the count_streaming metrics — _abs_dist_norm now coerces to float and guards divide-by-zero, and _parse_prediction_list tolerates malformed / non-list JSON. Scoring note: this corrects a bug where malformed predictions and zero-valued targets scored a perfect MRA (the old except: return 0.0 fed the "distance" side); they now score worst-case, so affected historical count_streaming numbers will drop.
• mmsi_video: Add a single-visual-input subset (mmsi_video_single_input) matching Cambrian-P's reported MMSI-Video protocol, and drop unused imports.
• tomato: Make the OpenAI import lazy (only needed for the disabled GPT parser) and stop hardcoding a 16-frame assumption in the prompt so the model's own num_frames governs the frame budget.
• videoevalpro: Guard the score aggregator against an unexpected task_type (previously a KeyError) and use identity comparison for the judge flag.

Validation

• py_compile, ruff check, black --check (line-length 240), and isort --profile black --check all pass on the changed files.
• YAML task configs verified; all referenced !function utilities exist.
• New erqa/tests/test_erqa_utils.py covers answer extraction, message interleaving, and aggregate accuracy.
• Offline verification of the egoschema metric change: TaskManager resolves egoschema_subset/egoschema_subset_mcppl with metric_list=['score'] and the full egoschema/egoschema_mcppl still with ['submission'], no network access.

Review pass (2026-07-06)

Self-review follow-up commit removes the submission metric from the two offline subset yamls (see egoschema bullet) — previously the README recommended the subset for offline eval while its submission metric still POSTed to the validation server and, after this PR's own retry hardening, would have raised on a no-egress node before the offline score was emitted.