Add lazy eval during conversion

[Blaizzy]

This PR implements lazy evaluation during the model loading and conversion process to improve memory efficiency. By utilizing MLX's lazy evaluation feature, we defer actual memory allocation until weights are needed.

Key changes:

• Added lazy=True parameter when fetching models from HuggingFace hub
• Memory allocation now happens on-demand rather than eagerly
• Works with both float16 (quantized) and user-specified dtype configurations

Benefits:

• Reduced peak memory usage during model conversion
• Smoother handling of large models, particularly when converting to float16
• Maintains compatibility with existing model loading workflows

Read more here: https://ml-explore.github.io/mlx/build/html/usage/lazy_evaluation.html