rollout notebook: fixups to benchmarks, plots, text

[aftersomemath]

Changes:

1. MJX benchmarks use jax.lax.scan and block_until_ready
2. Benchmark parameters updated to match RL use case
3. humanoid benchmarks are warmstarted by 2 seconds, humanoid100 by 4 seconds
4. benchmarks use skip_checks and preallocated threadpools
5. Improved advanced usage plots
6. Benchmark no longer includes model initialization in timed section
7. Update notebook text as needed

The benchmark results are qualitatively similar. Except for the N humanoid benchmark where the MJX performance is significantly improved.

It would be helpful if someone can double check the MJX benchmark.