Add embedding models support

🚀 Overview

This PR adds comprehensive embedding model support to Swama, enabling text embeddings generation using MLX-powered models for semantic search, similarity analysis, and other embedding-based applications.

✨ Key Features

• EmbeddingRunner: Core embedding inference engine with batch processing
• HTTP API: OpenAI-compatible /v1/embeddings endpoint
• Batch Processing: Efficient handling of multiple texts with automatic padding
• Usage Tracking: Token count reporting for monitoring

🛠️ Implementation

Core Components

• EmbeddingRunner.swift - Actor-based embedding generation
• EmbeddingsHandler.swift - HTTP API endpoint
• Integration with existing ModelManager

API Example

POST /v1/embeddings
{
  "input": ["Hello world", "Another text"],
  "model": "mlx-community/Qwen3-Embedding-4B-4bit-DWQ"
}

📦 Dependencies

• mlx.embeddings: New dependency for embedding inference
• MLX Swift: Updated to v0.25.4

✅ Quality Assurance

• Swift build passes
• SwiftFormat compliance
• Unit tests added
• Memory efficient implementation
• Thread-safe actor design

🔧 Technical Details

• Resolved SwiftFormat acronyms rule conflicts with external APIs
• Comprehensive error handling and validation
• OpenAI-compatible response format

🎯 Performance

• Batched inference for multiple texts
• Attention masking for variable-length sequences
• Efficient MLXArray operations

Breaking Changes: None - fully backward compatible

Testing: Unit and integration tests included