feat(embeddings): support offline FastEmbed startup

[bertheto]

Summary

• Adds opt-in FASTEMBED_LOCAL_FILES_ONLY=true for FastEmbed embedding and reranking model loading.
• Preserves default behavior for users who rely on first-run model downloads.
• Delays FastEmbed imports until adapter construction and passes local_files_only=True when offline mode is enabled.
• Wraps local-cache load failures with actionable guidance that includes the model and cache path.

Why

In corporate or offline environments, HuggingFace/GCS downloads may be blocked. Forgetful already supports FASTEMBED_CACHE_DIR, but if the cache is missing or incomplete, startup can still attempt remote model resolution. This gives deployments a deterministic cache-only mode that fails fast instead of attempting network access.

Validation

• uv run pytest tests/integration/test_fastembed_offline.py tests/integration/test_openai_embeddings_adapter.py tests/integration/test_http_reranker_adapter.py -> 26 passed.
• uv run python -m compileall ... on changed modules/tests -> passed.
• Empty-cache negative scenario with FASTEMBED_LOCAL_FILES_ONLY=true returns the new actionable RuntimeError.
• Offline MCP tool-list smoke with FASTEMBED_LOCAL_FILES_ONLY=true and RERANKING_ENABLED=false still exposes the 3 meta-tools.
• Full cache-present FastEmbed startup was not run locally because the configured cache exists but has no .onnx model files.

Compatibility

• Default remains FASTEMBED_LOCAL_FILES_ONLY=false.
• Existing Azure, Google, OpenAI, Ollama, and HTTP reranker paths are unchanged.
• This PR is independent from feat(mcp): add compact descriptor mode #43, which only addresses compact MCP descriptors.

[ScottRBK]

Please can we add a docker/.env.example entry for FASTEMBED_LOCAL_FILES_ONLY=false

aside from that LGTM.

[bertheto]

Thanks for the review! Added in c2d9e06 — FASTEMBED_LOCAL_FILES_ONLY=false is now in docker/.env.example right after the FastEmbed embedding block, with a short comment noting that the flag covers both the embedding model and the FastEmbed reranker (so contributors don't have to read docs/OFFLINE_SETUP.md to know it applies to the reranker too).

Ready for another look when you have a minute.