[BUG] NonTensorData.empty(names=...) ignores the names it is given #2501

NonTensorData.empty() and MetaData.empty() accept names, but the result never gets them: it keeps the old names, or has none when a batch_size is also given without names. TensorDict.empty(names=...) sets the names, so code that passes names to empty() on any tensor collection gets the wrong names back when the collection is a non-tensor entry.

from tensordict import NonTensorData, TensorDict

print(NonTensorData("x", batch_size=[2]).empty(names=["n"]).names)
nt = NonTensorData("x", batch_size=[2], names=["m"])
print(nt.empty(names=["n"]).names, nt.empty(batch_size=[2], names=["n"]).names)
print(TensorDict(batch_size=[2]).empty(names=["n"]).names)

[None]
['m'] ['m']
['n']

Expected: ['n'] in every case, as TensorDict.empty(names=["n"]) gives.

Cause: NonTensorDataBase.empty (tensordict/tensorclass.py:5527-5536) sets names = None when a batch_size is given without names, and otherwise always replaces names with self._maybe_names(), so a names argument is never used.

Possible fix: use names when it is given, else None with a new batch_size, else self._maybe_names(). Only calls that pass names change, and those return the wrong names today. Found while working on #2467.

tensordict main (eba7879), torch 2.14.1, Python 3.12.12.