Improve model loading: cache resolutions, share ModelManager, configurable quantization

[devin-ai-integration]

Summary

Addresses several model loading inefficiencies found during code review:

1. Cache _model_id() in TransformersRuntime — Previously _model_id() called ensure_model() on every invocation (2× per load in _load_generator and _load_embedder). Now caches the result in _resolved_model_id after the first call.

2. Share ModelManager instances across runtimes — Added get_model_manager(cache_dir) factory with double-checked locking that returns shared instances keyed by cache_dir. Previously each runtime constructor created its own ModelManager.

3. Cache GGUF file selection — Added _gguf_selection_cache dict to ModelManager so that once a repo's GGUF files are listed and the best quant selected, subsequent calls skip the list_repo_files() HF API call entirely.

4. Configurable quantization preference — Added preferred_quantization: str | None field to TransformerModelProfile. The llamacpp model support handler reads this (falling back to extra_options["preferred_quantization"]) and passes it to _select_gguf_file(), which prepends it to the preference order.

5. Fix pre-existing mypy error — Added llama_cpp.llama_chat_format to the mypy ignore_missing_imports overrides (the top-level llama_cpp entry didn't cover submodules).

Review & Testing Checklist for Human

• Verify the get_model_manager() shared-instance pattern is acceptable — it uses module-level state (_MANAGER_INSTANCES). If tests need isolation, ModelManager() can still be constructed directly.
• Verify _gguf_selection_cache keyed by "{repo_id}::{preferred_quantization}" is correct — if a user changes quantization preference between calls for the same repo, a new cache entry is created.
• Check that preferred_quantization on TransformerModelProfile being a frozen dataclass field with None default doesn't break any downstream serialization.
• Run the full test suite (pytest tests) — all 81 tests pass locally with ruff + mypy clean.

Notes

• The GLiNER PII backend already had process-level model caching (_MODEL_CACHE + _CACHE_LOCK). These changes bring similar patterns to the core model loading infrastructure.
• Direct ModelManager() construction still works for tests and one-off use; get_model_manager() is used by the runtime constructors and CLI.

Link to Devin session: https://app.devin.ai/sessions/f84a43ef798f44b3b1623ec8657b673c
Requested by: @donvito

---

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 4 additional findings.