[Feature] Lock tensordict when calling `TensorDict.share_memory_()`

Description

When a tensordict is shared in memory, all of its tensors have a storage that is changed accordingly.

We need to decide what happens when a tensor is added to the tensordict. Here are the various use-cases:

1. The tensordict is shared, the tensor is not, the key already exists, inplace
2. The tensordict is shared, the tensor is not, the key already exists, out-of-place
3. The tensordict is shared, the tensor is not, the key does not exist
4. The tensordict is on cuda (hence it is shared) and all of the above

Proposed solution

We could have a flag
share_memory_(lock=True) that ensures that no new key is added to the tensordict.
This would require the following changes:

• Ensure that lock_tensordict only prevents adding keys that are not present. Basically, right now locked tensordict prevent any kind of update. As this feature is not used anywhere in the code, we should simply ensure that locked tensordicts only accept in-pace updates (i.e. update of existing tensors without creation of new tensors in memory). td.set(..., inplace=True) would work iff the tensors already exists. td.set_ should not check lock status.
• implement the new flag for share_memory_. Not that tensordict on cuda are shared but do not need to be locked (i.e. we don't want every cuda-based tensordict to be locked). Only those on cpu should be locked by default.
• write tests for locked tensordicts in test_tensordict.py. Also write a test for locked tensordicts after calling share_memory_ with and without the flag being True.
• Implement similar changes for memmap_(lock=True) but this time also for cuda-based tensordicts.