feat: safetensors hyperparameter extraction with GGUF parity

[afogel]

Summary

• Extracts hyperparameters from safetensors repos by combining config.json (llama.cpp's find_hparam key fallback chains), tokenizer_config.json, and safetensors tensor headers via huggingface_hub.get_safetensors_metadata()
• Deduplicates config.json parsing into a canonical config_parsing.py module shared by both safetensors_metadata.py and extractor.py
• Safetensors takes precedence over GGUF in default_extractors() since safetensors is the original source format that GGUF is derived from

Changes

File	Purpose
src/models/config_parsing.py	Canonical HPARAM_KEYS + parse_config() — llama.cpp key fallback chains, VLM text_config merge
src/models/safetensors_metadata.py	SafetensorsModelInfo, map_to_metadata(), fetch_safetensors_metadata() — config + tokenizer + tensor headers
src/models/model_file_extractors.py	SafetensorsFileExtractor + reorder default_extractors() (safetensors first)
src/models/extractor.py	Wire _build_hyperparameters_from_config() into _try_config_extraction for the hyperparameter field
tests/test_safetensors_metadata.py	40 tests: config parsing, metadata mapping, tensor extraction, HF Hub integration, extractor wiring, precedence, fixture end-to-end
tests/test_model_file_extraction.py	Safetensors integration tests + fix pre-existing monkeypatch targets
tests/test_hyperparameter_wiring.py	Hyperparameter field flows through EnhancedExtractor pipeline
tests/fixtures/__init__.py	build_safetensors_fixture() — real safetensors binary for end-to-end tests
pyproject.toml	Add safetensors>=0.4.0 runtime dep, dev deps to [dependency-groups]

Design decisions

• config.json is format-agnostic: The same parse_config() works for safetensors, GGUF, pytorch — any HF repo with a config.json. Extracted to a shared module to avoid duplication.
• Header-only reading: Uses huggingface_hub.get_safetensors_metadata() (100KB range request) — never downloads full model weights. config.json and tokenizer_config.json are small JSON files fetched via hf_hub_download().
• GGUF parity: Safetensors extractor now produces equivalent metadata fields: model_type, typeOfModel, vocab_size, context_length, tokenizer_class, hyperparameter dict (including rope_dimension_count), plus safetensors_total_parameters.
• No changes to downstream: Output uses the same hyperparameter dict format that extractor.py and service.py already handle generically.

Test plan

• uv run pytest tests/test_safetensors_metadata.py -v — 40 tests
• uv run pytest tests/test_model_file_extraction.py -v — 17 tests (including new safetensors + precedence tests)
• uv run pytest tests/test_hyperparameter_wiring.py -v — 8 tests
• uv run pytest tests/test_gguf_metadata.py -v — 18 tests (no regressions)
• All 83 tests pass