update constants for `float`

[thowell]

mujoco has a number of constants that were tuned for double. we should determine good default values that work for float and update mujoco warp's relevant constants to these values.

• mjMINMU: mjMINMU (=1e-5) does not prevent nans when using float
• GJK tolerance: tolerance too small for float leading to failing test
• solver tolerance override

[adenzler-nvidia]

I hit a similar issue in #902 with a constant in GJK, there we have -1e-10 as a threshold which is too small for some float-based computations.

[thowell]

creating a list of constants to track in the pr description