autostart only affects new trials

[leoscholl]

the autostart feature -- which should really be renamed to random_start_time or something like that -- only affects the timing from the wait state to the target state, but not the timing in between hold penalty and target, for example. this means we get randomness only if the subject gets things correct.

maybe this was intended? either way it seems wrong.

[leoscholl]

ideally every timing parameter should be a distribution, possibly with 0 variance. but if the subject gets an incorrect response, nothing is resampled until the next wait state.

[leoscholl]

i think this is ok. other random parameters like random delay should also be sampled once per wait state so multiple attempts will always have the same (possibly difficult) delay settings.