Experience Buffers
- ReplayBuffer stores off-policy transitions and samples
them with replacement.
- RolloutBuffer stores one fixed-size on-policy rollout and
computes generalized advantage estimates.
Both allocate storage on the supplied linear-algebra backend. Observations and
actions passed to add() must therefore be compatible with that backend.