feat(models): Florence-2 end-to-end integration and real-checkpoint validation (sub of #850)

[inureyes]

Summary

The capstone that wires Florence-2 into mlxcel's model detection, registration, and generation paths, then validates the whole model on a real checkpoint across caption + OCR + detection modes. No new model math lives here; it connects the pieces from #852, #853, #854, and #855 into the runtime and proves the integrated deliverable.

Part of epic

#850

Dependencies

#852 (seq2seq engine + text core), #853 (DaViT backbone), #854 (fusion + weight loading), #855 (processor + task prompts + location tokens)

Upstream reference

• florence2.py: https://github.com/Blaizzy/mlx-vlm/blob/main/mlx_vlm/models/florence2/florence2.py
• config.py: https://github.com/Blaizzy/mlx-vlm/blob/main/mlx_vlm/models/florence2/config.py

Architecture notes

• This sub-issue contains no new model math. It connects the pieces from feat(models): Florence-2 BART-style seq2seq encoder-decoder engine + text core (sub of #850) #852 to feat(models): Florence-2 processor, task prompts, and location tokens (sub of #850) #855 into the runtime and proves the integrated model.

Implementation plan

• Add the "florence2" arm to src/models/detection.rs.
• Register in src/model_metadata.rs via for_each_model_registration!.
• Wire the seq2seq VLM decode path into generate / generate_vlm.
• Add the TP / distributed arch-string if applicable.
• Update docs/supported-models.md.
• Add _tests.rs unit tests (DaViT shapes, cross-attention, task-prompt parse).
• Validate on a real checkpoint (mlx-community/Florence-2-base-ft-bf16 or -4bit) across caption + OCR + detection modes via ./target/release/mlxcel generate -m <model> --image <img> -p "<task prompt>", confirming mlxcel list reports florence2.
• Follow docs/adding-models.md.

Acceptance criteria

• "florence2" arm added to src/models/detection.rs; registered in src/model_metadata.rs via for_each_model_registration!.
• seq2seq VLM decode path wired into generate / generate_vlm; mlxcel list reports florence2.
• TP / distributed arch-string added if applicable.
• docs/supported-models.md updated; _tests.rs unit tests added.
• Real checkpoint loads and produces correct output for at least caption + OCR + detection modes.
• Integration: the epic deliverable is a functionally integrated model reachable from the CLI and server, not standalone modules.

Effort

MEDIUM.

Note: this sub-issue is tracked by the checklist in #850.