This overhauls the Jax compiler and planner as follows:
action bounds are automatically computed but can be overridden in the config (this ensures backward compatibility)
nonlinear constraint is exposed by the jax compiler in the form of h(s, a) >= 0 where h is returned over batches and time step in the log provided by compile_rollouts
small improvements in numerical robustness in the jax planner
This overhauls the Jax compiler and planner as follows: