[bug-fix] Use float64 when converting np.ndarray to torch.tensor, cap Torch version to 1.7.x

[ervteng]

Proposed change(s)

PyTorch changed the behavior of tensor conversion - previously, if dtype was None it used the default tensor type. Now, if None it will use the type of the thing that is being converted.

This PR forces everything to use float32. In addition, we cap the Torch version to 1.7.X so that we won't break when Torch 1.8.0 comes out and breaks something.

Types of change(s)

• Bug fix
• New feature
• Code refactor
• Breaking change
• Documentation update
• Other (please describe)

Checklist

• Added tests that prove my fix is effective or that my feature works
• Updated the changelog (if applicable)
• Updated the documentation (if applicable)
• Updated the migration guide (if applicable)

Other comments

[chriselion]

Not directly related to this change, but can we put a link to the windows instructions (either to our docs or external docs) as a comment here?

[chriselion]

Looks good. If the float32 restriction is a pain and doesn't provide any performance benefit, we can remove the checks for it.