[mlx_lm] Expose 'strict' parameter in load() function

[zyguy]

Motivation: Currently, the high-level mlx_lm.load() function does not expose the strict parameter. This makes it impossible to pass strict=False when loading certain converted models or "abliterated" models that contain harmless orphaned weights.

The underlying load_model() function already supports the strict argument. This PR simply surfaces it to the main load() entrypoint, maintaining True as the default to ensure full backward compatibility. This will allow dependent projects like oMLX to load a wider variety of fine-tuned or abliterated models without crashing.