Restore #460 seed-removal semantics after #339

[janosh]

• restore intentional Introduce PRNG to SimState and add reproducibility docs. #460 API behavior where NVE/NVT/NPT init functions do not take a seed kwarg and instead use state.rng
• switch integrator init paths back to initialize_momenta(..., generator=state.rng) and remove the reintroduced calculate_momenta helper
• update affected examples/tutorials and test helper to follow state-bound RNG seeding (state.rng = ...)

Context

These are surgical restorations of intentional changes from #460 that were unintentionally lost during the large ty-fixing merge in #339.

@CompRhys sorry about this oversight on my side while polishing the #339 LLM type fixes. I think this surgical restore might be easier to review than the broader revert path in #476.

[thomasloux]

Overall looks good to me, but I haven't checked that it covered all cases of seed.

[thomasloux]

@janosh actually I've added count_degrees_of_freedom when I added constraints. But to be honest, this function is not used in torch_sim and actually a user should use state.constraints = my_constraints and then use state.get_degrees_of_freedom. I'm preparing a PR to remove this function.

[thomasloux]

If it happens that some people use this function, actually the fix is change the type annotation. It should return a torch.Tensor and not a int. That's actually the error on my side + changing the check for non negativity.

[CompRhys]

@CompRhys sorry about this oversight on my side while polishing the #339 LLM type fixes.

No harm :) as long as we end up eventually correct when we make a release that's all that matters

[CompRhys]

Will merge this and let #479 resolve itself separately

[CompRhys]

lgtm