Better quantized support investigation

[polvalente]

PR #96 added initial support for quantization and quantized operations.
However, the example claims that >100 tok/s is achievable.

We need to investigate the example for any improvements we can make, and possibly import it as a test or benchmark

Running 8B parameter models requires 16GB+ at fp16. With 4-bit quantization, the same model fits in ~5GB, enabling inference on consumer Macs. This work is part of a broader effort to bring production LLM inference to the Elixir ecosystem:

Repository	Purpose
bobby_posts	Pure Elixir Qwen3-8B inference (135 tok/s)
bobby_posts_adapters	LoRA fine-tuning for personalized generation
bumblebee_quantized	Quantized model loading for Bumblebee
safetensors_ex	MLX 4-bit safetensors format support