[BUG] MemmapTensor indexing does not behave like indexed torch.Tensor

Describe the bug

When indexing a MemmapTensor, we would expect that the indexed tensor shares the memory of the parent tensor whenever this happens with torch.Tensor.

To Reproduce

from torchrl.data import MemmapTensor
import torch
x = torch.ones(3)
y = MemmapTensor(x)

x0 = x[0]
x0.copy_(torch.zeros([]))


y0 = y[0]
y0.copy_(torch.zeros([]))

print(x, y.clone())

results in

tensor([0., 1., 1.]) tensor([1., 1., 1.])

However this works:

x[None].fill_(0.0)
y[None].fill_(0.0)
print(x, y.clone())

tensor([0., 0., 0.]) tensor([0., 0., 0.])

Reason and Possible fixes

• We could use torch.from_numpy but this won't work on cuda.
• We can create another (sub-)MemmapTensor that shares the same file pointer but has a stored index. The problem then becomes: what happens if the parent MemmapTensor gets out of scope? The file will likely be deleted, and the sub-MemmapTensor will be orphan. Also, working with the sub-MemmapTensor will come with some overhead as every time we access its data we must index the array on disk.