end simulation correction when simulation has callbacks

[rnwatanabe]

Sometimes with long simulations with callbacks, the simulation never ends. Here is a small correction in the run_until() method in the control.py file that eliminates this issue.

[apdavison]

Thanks for reporting the bug. I think that for backends that round run_until() to the timestep grid (NEURON, NEST, Brian), if time_point doesn't land exactly on the dt grid, t gets stuck at the last grid point below it, the guard simulator.state.t + 1e-9 < time_point never becomes false, and the loop spins forever. That matches the "long simulations with callbacks never end" symptom, since time_point tends to drift off exact grid alignment through floating-point accumulation.

The problem with your proposed solution is that it stops the callback loop up to a full timestep early, even in the normal on-grid case, so run_until ends short of time_point and any callbacks scheduled in that final timestep are dropped.

Could we instead use the same half-timestep tolerance already applied a few lines above (the -dt/2.0 check when detecting a past time_point)?

  while time_point - simulator.state.t > simulator.state.dt / 2.0:

This terminates the off-grid case (it stops once within half a timestep of the target, which is the grid-rounding tolerance) without stopping short or dropping the final-timestep callbacks.

It'd also be worth adding a regression test that reproduces the hang. The existing test_callbacks doesn't catch it because it runs on the mock backend (whose run_until() sets t exactly, with no rounding) and uses callback steps much larger than dt, so it'd need a system test on one of the rounding backends, or an off-grid time_point against a backend that rounds.