[Feature Request] Transformers for first-class support in torchRL

I'm proposing TransformerModule: the transformer parallel of the LSTMModule, built on the same four aspects. Collection runs one step at a time against a fixed-shape KV cache carried per env; training processes full windows under a block-diagonal causal mask so attention never crosses episode boundaries. The load-bearing invariant, enforced by a parity test, is that the two paths produce identical outputs. One deliberate rule throughout: the KV cache is inference state, not experience. It lives in the rollout, never enters replay, and training recomputes state under current weights.

I am self-assigning this feature to myself. I have some experiments running for this