[BUG] Improvements on MemmapTensor

[vmoens]

Describe the bug

We need to do some woodwork on MemmapTensor:

• Remove the ownership and replace it by a check over the number of times the array is being opened (for temporary files only)
• Solve a bug where calling MemmapTensor(memmap_tensor[idx]) returns the wrong view of the MemmapTensor [Feature] apply across multiple TDs #193
• Improve MemmapTensor creation args, including: creating a Memmap on a specific location ([Feature] More arguments for MemmapTensor construction #189)
• Improve the way we copy a Memmap from one place to another, if we want and if we don't want to change the filename
• simplify the logic of MemmapTensor so that extracting a tensor is less confusing (i.e. check as_tensor, _tensor, _load_item and how they interact).
• the device attribute is confusing, as the data is never really on another device but gets loaded there if needed. The usage is that one can put the device attr as a destination, then move the Memmap from process to process and when accessing the data it'll be on the right device. However, this may give the false illusion that the content of the tensor is stored on the device...

cc @tcbegley

[tcbegley]

• AttributeError on Tensor attributes when creating one MemmapTensor from another. ([BugFix] MemmapTensor from MemmapTensor __dir__ issue #192 <- needs to be adapted to recently merged [Feature] apply across multiple TDs #193)

I'll also update #189, though the desired no-op behaviour there is blocked by #192.