Lie group normalization on LieTensor

[luisenp]

Now that @fantaosha implemented normalization operator in #506, one question is what how do we want to use this in LieTensor. Some ideas:

1. All tensor inputs are checked for group properties. If broken, emit a warning and normalize.
2. All tensor inputs are normalized without checking.
3. Constructor has an input flag that's something like force_normalize.
4. Constructor has an input flag that's something like maybe_normalize (check then normalize).
5. Add a context that global normalization behavior. Defaults to True. If False no checks/normalization.

Something else? Looking for input from @mhmukadam @rmurai0610 @joeaortiz @dongjing3309 @maurimo @ddetone

[mhmukadam]

What tradeoff are we considering, for instance speed vs ease of use (too many warning, confusing api).

5 is a nice option for users to do this in chucks of code using with. Related to this we could consider a fully global context like torch.set_default_tensor_type().

3/4 seems useful if users want to do this for any specific variable. For 4 check_then_normalize would be clearer.