Infer correct dtype for scipy.spatial.transform.Rotation

[Jiayuan-Gu]

Is your feature request related to a problem? Please describe.
When calling apply method on an nx3 array, it will become np.float64 anyway.

Describe the solution you'd like
Is it reasonable to infer the output dtype according to the input?

Describe alternatives you've considered

Additional context (e.g. screenshots)

[pmla]

@nmayorov what is your opinion on this? It is quite easy to implement for the `float32' type, but it might also be surprising for a user.