Inconsistent direction inputs expected for `SHEncoding.forward()` when `implementation = torch/tcnn`

[wengflow]

When implementation = torch, the in_tensor input of SHEncoding.forward() is expected to be unit vectors normalized between [-1, 1], as expected by components_from_spherical_harmonics(). However, when implementation = tcnn, in_tensor is expected to be vectors normalized between [0, 1], as expected by tcnn.Encoding. This inconsistency also propagates to NerfactoField and GenerfactoField.