This PR introduces a fused QKV (Query-Key-Value) implementation in the attention module for BitNet-1.58-2B on MLX. The fusion significantly improves prompt and generation speed by around 10%.

Key Changes:

• Added support for fused QKV projection in the attention layer.
• Updated model forward pass to conditionally use the fused path when enabled.

Benchmarked performance improvements (M3 Max)"

• Prompt Processing: ↑ from 128.77 to 139.94 tokens/sec
• Generation Speed: ↑ from 67.05 to 73.12 tokens/sec
• MLX Fused QKV vs BitNet 4T = 27.6% faster generation, 137% faster prompt