[BugFix] Reject a names list whose length is not batch_dims

[YeonwooSung]

Fixes #2600

TensorDictBase._set_names counted None entries and, when that count equalled batch_dims, cleared the names and returned before checking that len(value) == batch_dims. A list one entry too long was therefore accepted whenever it held exactly batch_dims Nones, and every name was dropped:

TensorDict(a=torch.zeros(3), batch_size=[3], names=["s", None]).names  # was [None], no error
# names=["s", "t"] already raised ValueError

The same shortcut ran for a lazy stack. s.names = ["s", "t", None] on members of batch [3] forwards ["t", None] to each member, which has exactly one None. The member cleared its names and "t" was dropped. Checking the length in the base method rejects that list before any name is written on the member, so the members are not left half-updated.

The length check now comes first and raises the existing ValueError (the length of the dimension names must equate the tensordict batch_dims attribute). Names that were already set are left unchanged. An all-None list of the right length still clears names. A right-length mix of strings and Nones still assigns.

This is a silent data loss, so it raises now. It is not a scheduled default change.