Support embedding models

Summary

GenieX currently only supports generative models (text-generation / VLM) — there's no way to run an embedding model to get vectors. Would be great to add embedding support.

Tried LiquidAI/LFM2.5-Embedding-350M (Q4_0 GGUF, lfm2 arch) on WoS ARM64 (X Elite), CLI v0.3.13:

• pull detects it as llm
• infer forces generation → degenerate output (password password password...)
• serve's POST /v1/embeddings → 404 (server only has /v1/models and chat completions)

The same GGUF returns proper 1024-dim vectors via upstream llama.cpp's llama-embedding, and cosine similarity behaves correctly — so the model is fine and the lfm2 arch is already embeddable; GenieX just doesn't expose the path.

The underlying llama_cpp runtime should already have this (llama.cpp's llama_get_embeddings). Would be nice to:

1. Detect embedding models on pull
2. Add a /v1/embeddings endpoint (and/or a CLI/SDK embed API)

On-device RAG / semantic search is a common use case. Thanks!