Remove set_default_dtype and get_default_dtype

[PicoCentauri]

Fixes #148

This PR removes all calls of set_default_dtype and get_default_dtype. For now I am using the .to() method to move the model to the correct dtype. But is there also an option to initialize the model in the first place in the correct dtype?

The reveiled some issues that are also fixed within this PR

1. if a user changes the dtype for continue training the soap-bpnn would not work. Now the dtype of the model is set to the one of the dataset.
2. similar for eval. There we adjust the loading dtype to the targets to the dtype the exported model

Additional changes

• changed the default precision to 32 to match the torch one. We can also leave this at 64-bit?
• replaced assert torch.allclose by the more verbose torch.testing.assert_close

TODO

• add an eval test for 64-bit and 32-bit exported models
• add __architecture_capabilities__ to each architecture indicating their supported types
• add dtype check to check_datasets function

---

📚 Documentation preview 📚: https://metatensor-models--155.org.readthedocs.build/en/155/

[frostedoyster]

Thanks for the changes

[frostedoyster]

Uhhh are we sure we want this to be the default behavior? People have had issues with float32 MD

[PicoCentauri]

I just changed this to be consistent with torch. Which people had issues? The examples seems to be good and usually problems in MD appear if the forces are not summed up in double precision but as a base precision I think a normal float should be fine.

[frostedoyster]

Ok, we can discuss this together with other people. I still think float64 is a safer default

[PicoCentauri]

I am using the power of omegaconf's custom resolvers to obtain the preferred dtype for each architecture. When we are happy with this I will do the same for the device (but in a different PR)

[frostedoyster]

Thanks!