[MLX] [bugfix] Preserve dtype of array when converting to torch

[neilmehta24]

We have noticing the following error with a recent version of outlines when used with MLX:

TypeError: argument 'token_id': 'float' object cannot be interpreted as an integer

At:
  /.../outlines_core/fsm/guide.py(294): get_next_state
  /.../outlines/processors/structured.py(101): process_logits
  /.../outlines/processors/base_logits_processor.py(90): __call__

The issue is that the MLX array of tokens, which are integers, are being force-converted to floats, even though outlines expects an integer array. This is because all MLX arrays are being converted to float32, even when it's not necessarily appropriate, like in this case. Looking at the commented link, the advice was to convert to float32 only for bfloat16, because numpy does not support bfloat16. Now the MLX _to_torch implementation matches the other array libraries, none of the other libraries are being force-casted to float

[neilmehta24]

@rlouf The CI configuration prevents me from reaching the coverage requirements. From the CI logs:

outlines/processors/base_logits_processor.py (0.0%): Missing lines 113-115

But note that the CI doesn't install or run any MLX code, so to my understanding it's impossible to improve the CI coverage report when adding MLX-specific logic.

However, I did check locally that there isn't a regression in coverage, and that the tests are passing

Coverage:

Test:

So this should be good to go, please let me know what you think.