[BUG] `Head._task_weights = defaultdict()` silently behaves like a plain dict

Bug description

PR #802 (commit ab7207cf) changed self._task_weights = defaultdict(lambda: 1.0) to self._task_weights = defaultdict() in transformers4rec/torch/model/base.py:

https://github.com/NVIDIA-Merlin/Transformers4Rec/blob/8bf122f5dcb39feecfc6dabde734d79c2d1c4380/transformers4rec/torch/model/base.py#L272

self._task_weights = defaultdict()

defaultdict with no factory argument has the same behavior as a plain dict — missing keys raise KeyError. The previous defaultdict(lambda: 1.0) returned the documented default 1.0 for tasks without an explicit weight.

Only one usage was patched to .get(name, 1.0). Direct indexing sites (e.g. future code, or external subclasses that read head._task_weights[task_name]) will now raise KeyError where they previously returned 1.0.

Steps/Code to reproduce bug

from collections import defaultdict

old = defaultdict(lambda: 1.0)
new = defaultdict()

old["anything"]   # -> 1.0
new["anything"]   # KeyError: 'anything'

Expected behavior

Either restore the lambda: 1.0 factory, so defaultdict conveys its documented meaning, or switch to plain dict() with consistent .get(name, 1.0) access at every call site. Using defaultdict() is misleading because the name promises a default factory it no longer has.

Environment details

• Transformers4Rec: main @ 8bf122f5 (regression from PR #802 / commit ab7207cf)

Additional context

Preferred fix — restore the factory (closest to original behavior):

self._task_weights = defaultdict(lambda: 1.0)

Happy to send a one-line PR.

Open transformers4rec/torch/model/base.py and inspect the _task_weights initialization against the prior factory behavior described in the issue. Verify the reproduction and relevant existing tests; done means missing task names return 1.0 while explicit weights continue to work.