Make unroll_length schedulable

[QuantuMope]

This PR allows for a scheduled unroll length if we are running synced off-policy RL training:

1. async_unroll=False
2. whole_replay_buffer_training=False

It also allows for a scheduled value of 0, which in turn skips unrolling to train from the replay buffer.

This allows us to "simulate" very diverse training strategies. E.g.,

unroll_length = StepScheduler("iterations", [
    # unroll to collect an "offline" dataset
    (1, int(initial_collect_steps / num_para_envs)),        
    # perform offline training iterations. no unroll 
    (offline_training_iters, 0),     
    # continue with online RL                                    
    (offline_training_iters + 1, desired_unroll_length)])

Codex cleverly makes a minimal change with full backward compatibility by adding the following code

    @property
    def unroll_length(self):
        return self._unroll_length()

    @unroll_length.setter
    def unroll_length(self, value):
        self._unroll_length = as_scheduler(value)

[Haichao-Zhang]

ConstantScheduler --> should check against a base class, e.g. Scheduler?

[QuantuMope]

We need to check against ConstantScheduler here because a scalar input will be converted to one before this check due to the setter function on line 479.

Any non-constant scheduler should then raise an error if we're doing on-policy or async unroll.

[QuantuMope]

Hey Haichao, responded to your comment. Let me know if I misunderstood it.

[QuantuMope]

We need to check against ConstantScheduler here because a scalar input will be converted to one before this check due to the setter function on line 479.

Any non-constant scheduler should then raise an error if we're doing on-policy or async unroll.