Add deterministic memory cleanup across inference paths:
- Detach logits after Eval to release graph references
- Free intermediate arrays in attention (gemma3, qwen3)
- Add cache Detach helper for KV cache cleanup after generation
- New detach.cpp/go CGO bindings for mlx_array_detach
Reduces 4B model memory from 78GB to ~17GB (vs 2.4GB mlx-lm baseline).
Native Metal memory management still trails Python refcounting but is
now viable for 1B models.
Co-Authored-By: Virgil <virgil@lethean.io>
