Add MNEMON_EMBED_DIMENSIONS for Matryoshka dimension truncation

[achinth-b]

Closes #4

Support Matryoshka Representation Learning by passing the dimensions parameter to Ollama's /api/embed endpoint. When MNEMON_EMBED_DIMENSIONS is set (e.g., 256), Ollama truncates and re-normalizes the embedding vector, giving faster similarity search with minimal quality loss on MRL-trained models like nomic-embed-text.

Backward compatible: when unset, behavior is identical to before.

What

• Added dims field to the Client struct (0 = use native dimensions)
• NewClient() reads the MNEMON_EMBED_DIMENSIONS env var (parsed as int, ignores invalid/negative values)
• Added Dimensions field to embedRequest with json:"dimensions,omitempty" so it's only included when set
• Embed() conditionally populates req.Dimensions when dims > 0
• Updated CHANGELOG.md under [Unreleased]
• Updated docs/USAGE.md Configuration table with the new env var

Why

Here is the issue which describes the problem.

Matryoshka-trained models encode the most important semantic signal in the first N dimensions. Using 256 dims instead of 768 gives ~95% retrieval quality at 3× less storage and faster cosine similarity. This is especially impactful for large knowledge bases where similarity search dominates recall latency.

Verified locally

Tested against Ollama nomic-embed-text on macOS:

# Without dimensions → 768-dim vector
curl -s http://localhost:11434/api/embed \
  -d '{"model":"nomic-embed-text","input":"test query"}' 
# → Vector length: 768

# With dimensions: 256 → 256-dim vector
curl -s http://localhost:11434/api/embed \
  -d '{"model":"nomic-embed-text","input":"test query","dimensions":256}'
# → Vector length: 256

Warning

Re-indexing caveat
If users change dimensions on an existing store, old embeddings (768-dim) and new embeddings (256-dim) will have incompatible vector sizes. For this PR, the env var only affects newly generated embeddings. A follow-up could add a mnemon embed --reindex command.

[Grivn]

LGTM. Thanks for the clean, focused change.