Plans for v0.3.0

[vmoens]

The v0.3.0 release will bring the following features to TensorDIct:

• Better separation between base class and TensorDict as in [RFC] Tensordict integration pytorch#112441
• Faster operations on shape (split, unbind, etc) by pre-computing the batch-size for each.
  The issue here is that for these operations, we simply index the tensordict and return the indexed version which require to dynamically compute the batch-size. By pre-computing the batch-size, we can achieve a higher throughput in all cases.
• Faster iteration over leaves and nested TDs: we will be storing the leaves in tensordict._leaves and the nested tds in tensordict._nested. That way, we can avoid a lot of instance checks, making calls to items much faster (and better sorted). On a typical, highly nested apply call with a simple operation, isinstance accounts for 5-10% of runtime, which is roughly equal to the overhead when compared to pytree. We therefore expect this change to make apply be at par with tree_map.
• MemoryMappedTensor instead of MemmapTensor [Refactor] MemoryMappedTensor #541 which will speed up memory-map accesses.
• Progressive switch to plain stacking, viewing and permutations: we would like for tensordict ops to return plain TensorDict instances by default unless the user specifically asks for a lazy class. This change will be progeressive, with warnings in v0.3 and complete switch in v0.4. Lazy classes can be surprising as they introduce overhead where it isn't expected (eg, get).

Stretch goals:

• torch.compile compatibility.