fix: reconcile task registry into a single source of truth (#724)

[timenick]

Fixes #724.

Problem

loader/task.py held two hand-maintained task tables, created by different PRs from different authorities, that silently drifted apart:

• TASK_ABBREV (PR feat: add winml serve & winml run with MCP support #354) — canonical→abbreviation map for cache keys; intentionally broad (full HF taxonomy).
• KNOWN_TASKS + validate_task (PR fix(inspect): make --list-tasks return in <500 ms (#544) #717) — a hand-copied snapshot of optimum's exportable tasks for a fast winml inspect --list-tasks, never cross-checked against TASK_ABBREV.

PR #717's validate_task gate consults only KNOWN_TASKS, so the pre-existing divergence became a user-visible bug: 13 task names that lived only in TASK_ABBREV were rejected — including ones WinML genuinely supports through a third independent table, inference/tasks.py (TaskInputSpec). Root cause: no single source of truth for the canonical task set.

Fix

Both public views now derive from one private registry, so they can no longer desync:

• _TASK_REGISTRY: dict[str, str | None] — the single authoritative table of canonical task → abbreviation (None keeps the existing 8-char truncation fallback).
• KNOWN_TASKS = frozenset(_TASK_REGISTRY) and TASK_ABBREV (non-None abbrevs + 5 collapsed aliases) are both derived.

Audit of the 13 divergent names:

• 5 wired → added to the canonical set: video-classification, keypoint-matching, table-question-answering, zero-shot-audio-classification, any-to-any — each has a TaskInputSpec, a composite-model registration, and/or an e2e testset tag.
• 8 stale → dropped: instance-segmentation, universal-segmentation, masked-image-modeling, text-encoding, audio-tokenization, multimodal-lm, backbone, time-series-prediction — unreferenced anywhere outside the old table; masked-image-modeling / time-series-prediction were long-form duplicates of the existing masked-im / time-series-forecasting.
• 5 collapsed aliases (pretraining, sequence-classification, summarization, translation, zero-shot-classification) stay out of KNOWN_TASKS but keep their abbreviations, because some callers use the alias name directly as the resolved task (composite models register summarization/translation; inference.tasks defines a spec for zero-shot-classification).

Zero cache churn: every abbreviation still in use is byte-for-byte unchanged. Net: KNOWN_TASKS 35→40, TASK_ABBREV 47→39, and no abbreviation value modified — so no existing cache key changes.

Drift guards added

tests/unit/loader/test_known_tasks.py locks the invariant going forward:

• KNOWN_TASKS and TASK_ABBREV must derive from _TASK_REGISTRY; abbreviations must be unique (the serve reverse-decode map stays lossless).
• The 5 wired tasks must pass both validate_task and the click --task callback.
• The 8 dropped tasks must stay out of every view.
• Every inference.tasks TaskInputSpec key must be a known task — the cross-source check that would have caught this bug.