Missing functionality around types/dtypes

[davidkoski]

As observed in:

• Add Qwen 2.5 VL mlx-swift-examples#222

Perhaps DType should have finfo() capabilities:

• https://numpy.org/devdocs/reference/generated/numpy.finfo.html
• min, max, etc.

There is inconsistency in the creation methods for MLXArray. Some take dtype: and some take only type: (swift type). Further inconsistencies between free functions (e.g. zeros) and static functions (MLXArray.zeros)

[davidkoski]

finfo isn't that easy -- the values all have types and that won't be dynamic in DType itself.

[awni]

finfo isn't that easy -- the values all have types and that won't be dynamic in DType itself.

I'm not sure exactly what you're saying. But the way it works in C++ (and Python) is we just make the finfo struct use double for everything. That way min/max can be represented in finfo for all the types.

[davidkoski]

I could do that but there is no implicit coercion so you end up with a bunch of Float(finfo.max) potentially. But this is probably better than nothing!