Better dtype handling

[frostedoyster]

At the moment, we have that all the readers have a dtype argument with the torch.float64 default. However, with our current global dtype model, this does not make much sense and it makes maintaining the tests quite difficult at times (see #146). I would propose to remove the dtype argument and always use the global dtype, so as to make the logic consistent with our current dtype handling. I do not see it breaking our code in any way, because we always use the global dtype.
This could be changed in the future.
@PicoCentauri

[PicoCentauri]

Personally, I don't like the global dtype option. I know that you will argue that this standard and everybody is doing it, but I think this is an antipattern. Using a global variables makes code very implicit and very hard to debug if something goes wrong. I don't know what you saw in #146 but this seemed unrelated to the dtype.

I would propose to remove the dtype argument and always use the global dtype, so as to make the logic consistent with our current dtype handling.

I would rather say we remove the set global dtype otions and models should use the dtype they get from the dataset. The current implementation in mixed and was a compromise we did.

[Luthaf]

I agree with @PicoCentauri: if we allow setting global dtype, this mean that importing architecture_1 changes the behavior of architecture_2, which I would rather not have.

We can add a lint rule to forbid the use of torch.set_default_dtype() and torch.get_default_dtype() in metatensor-models, but we should have a review point that these are not used in external dependencies either.

[PicoCentauri]

I will make the dtypes explicit everywhere. @Luthaf do you know how to do such a lint rule?

[Luthaf]

Writing a flake8 plugin (https://flake8.pycqa.org/en/latest/plugin-development/index.html) would be one solution, or we could use a grep-based linting script.