fix(task): make detect_task architecture-aware for multi-task model types

[timenick]

What

detect_task short-circuited to the first (model_type, task) key in MODEL_CLASS_MAPPING, so every encoder-decoder model type (bart, t5, marian, mu2, qwen3, blip, vision-encoder-decoder) auto-detected as feature-extraction regardless of the architecture head — and disagreed with the already-head-aware config / build path (which uses _detect_task_and_class_from_config).

Change

Short-circuit step 1 only when a model_type maps to exactly one real (non-None) task. With multiple distinct tasks, fall through to architecture-aware detection. A (model_type, None) default-class sentinel no longer forces a fall-through, so single-real-task types like sam resolve to their real task (mask-generation).

Result — inspect now agrees with config / build

model	before (inspect)	after (inspect)	config / build
facebook/bart-large-mnli	feature-extraction	text-classification	text-classification
facebook/bart-large-cnn	feature-extraction	fill-mask*	fill-mask*
facebook/bart-base	feature-extraction	feature-extraction	feature-extraction
facebook/sam-vit-base	feature-extraction	mask-generation	mask-generation
clip / siglip	feature-extraction	feature-extraction	feature-extraction

* fill-mask for seq2seq generation models is still semantically wrong (optimum maps BartForConditionalGeneration to fill-mask). This PR only makes inspect consistent with config / build. Correcting *ForConditionalGeneration -> text2text-generation — which changes build artifacts and needs a text2text-generation export path — is Step 2, tracked in #838.

Tests

• Unit: ambiguous multi-task type falls through to architecture-aware detection; a single-real-task type with a None sentinel (sam) resolves to its real task; a single-entry type still short-circuits.
• Integration (test_task_consistency): added bart-large-mnli -> text-classification and sam-vit-base -> mask-generation.

Addresses #838 (Step 1 of 2).

[DingmaomaoBJTU]

Overall this is a correct and clean fix. The core logic is sound: collect only the non-None task entries for a given model_type, and short-circuit only when there is exactly one — otherwise fall through to the architecture-aware TasksManager path. Three issues worth addressing below.

[DingmaomaoBJTU]

3 of 4 previously raised issues are resolved. One remains open (see inline comment).

• Type safety (distinct_tasks: set[str]) ✅ Fixed.
• sam2 unit test (test_detect_task_falls_through_for_multi_task_model_type_sam2) ✅ Added.
• Network marker (pytestmark = pytest.mark.network) ✅ Applied module-wide with a helpful comment.
• Performance (O(n) scan) — Still open; see inline comment. Not a blocker given the current mapping size, but worth tracking.