Precompute inverse square root of impratio

[adenzler-nvidia]

I don't need this change in 100%. I just came across this, tried it and I can see a slight edge in benchmarks/kernel timings. Nothing compared to the bigger things we still have to optimize though.

That being said, apart from the interface change I think this has only benefits, as it removes runtime square root instructions in some kernel that run very often (linesearch) and sqrt throughput is generally lower than multiplication. My intuition and a very quick experiment tell me that a non-default impratio that doesn't have a trivial solution to sqrt makes these gains more visible.

Numbers for humanoid on an RTX Pro 6000 Blackwell.

impratio == 1:

main:

Summary for 8192 parallel rollouts

Total JIT time: 0.31 s
Total simulation time: 2.29 s
Total steps per second: 3,580,815
Total realtime factor: 17,904.08 x
Total time per step: 279.27 ns
Total converged worlds: 8192 / 8192

this pr:

Summary for 8192 parallel rollouts

Total JIT time: 0.31 s
Total simulation time: 2.28 s
Total steps per second: 3,591,947
Total realtime factor: 17,959.73 x
Total time per step: 278.40 ns
Total converged worlds: 8192 / 8192

Impratio == 7:

main:

Summary for 8192 parallel rollouts

Total JIT time: 0.32 s
Total simulation time: 2.62 s
Total steps per second: 3,128,554
Total realtime factor: 15,642.77 x
Total time per step: 319.64 ns
Total converged worlds: 8192 / 8192

this pr:

Summary for 8192 parallel rollouts

Total JIT time: 0.30 s
Total simulation time: 2.60 s
Total steps per second: 3,145,472
Total realtime factor: 15,727.36 x
Total time per step: 317.92 ns
Total converged worlds: 8192 / 8192