feat: per-stage model routing for the extraction pipeline

[miguelgfierro]

What

Every LLM pipeline stage can now run on its own model, configured via .env / environment variables. Stages left unconfigured fall back to the global FLYDOCS_MODEL (or the request's options.model when the caller sets one).

Stage	Variable	Default in env_template
splitter	FLYDOCS_SPLITTER_MODEL	anthropic:claude-sonnet-4-6
classifier	FLYDOCS_CLASSIFIER_MODEL	anthropic:claude-sonnet-4-6
extract	FLYDOCS_EXTRACT_MODEL	anthropic:claude-sonnet-4-6
visual_authenticity	FLYDOCS_VISUAL_AUTHENTICITY_MODEL	anthropic:claude-sonnet-4-6
content_authenticity	FLYDOCS_CONTENT_AUTHENTICITY_MODEL	anthropic:claude-sonnet-4-6
judge	FLYDOCS_JUDGE_MODEL	anthropic:claude-opus-4-8
rules	FLYDOCS_RULE_ENGINE_MODEL	anthropic:claude-sonnet-4-6
transform	FLYDOCS_TRANSFORM_MODEL	anthropic:claude-sonnet-4-6
bbox LLM matcher	FLYDOCS_BBOX_MATCHER_MODEL	anthropic:claude-sonnet-4-6

Resolution precedence (per stage)

FLYDOCS_<STAGE>_MODEL > options.model (request) > FLYDOCS_MODEL.

A pinned stage wins over options.model so operator stage-tuning survives per-request overrides.

How

• IDPSettings: nine new optional fields (None = inherit the global model).
• PipelineOrchestrator: a _stage_models() resolver computes the per-stage map once per request, stashed as ctx.metadata["stage_models"]; the seven per-call sites read their stage entry instead of the shared model_id.
• transform and the bbox LLM matcher receive their model at bean construction (core/configuration.py) since those services are not passed a model per pipeline call.
• Escalation is untouched: FLYDOCS_ESCALATION_MODEL / options.escalation.model still drives the judge-triggered re-run.
• Cost telemetry needs no change: pipeline.usage.by_model already splits usage per model id.

Tests

• 7 new unit tests (tests/unit/test_per_stage_models.py): settings defaults + env binding, resolver fallback + precedence, end-to-end orchestrator routing (extract + judge receive their pinned models), and bean wiring for matcher/transformer.
• Full unit suite: 305 passed. ruff clean; no new pyright errors (12 pre-existing on base).