Revisit Lie group consistency checks

[luisenp]

We have observed some 64-bit SE2 tensors with cosines slightly larger than ones sneaking through after application of retraction maps. We should consider making the Lie group checks stronger to prevent issues like this. Another option would be to normalize all exp_map outputs (e.g., here).

[mhmukadam]

Is this happening in the new Lie groups too?

[mhmukadam]

Flagging unresolved comments from #528 that need to be resolved with this issue.

• need greater eps for the log and exp derivative computation
• might need to add so3_derivative_near_zero_eps_floatxx

[luisenp]

Closed by #544