feat(inference): Implement Q3_K dequantization in aprender-serve

[noahgift]

Description

The aprender-serve inference engine currently lacks support for Q3_K (GGML type 11) dequantization. When attempting to load a Q3_K model (e.g., qwen2.5-7b-instruct-q3_k_m.gguf), inference crashes with:

Inference failed: Operation 'get_tensor_f32' not supported: Unsupported quantization type: 11

Acceptance Criteria

• Implement dequantize_q3_k in crates/aprender-serve/src/quantize/ (both standard and SIMD/parallel variants if applicable).
• Hook the implementation into the get_tensor_f32 match arm in crates/aprender-serve/src/gguf/metadata.rs.
• Add coverage tests for Q3_K tensors.