Custom callable is ignored in retrieval metric aggregation

[fschlatt]

🐛 Bug

When passing a custom aggregation to any RetrievalMetric, the callable is ignored. There is a bug in the retrieval aggregation function that defaults to using the maximum of the values by default

torchmetrics/src/torchmetrics/retrieval/base.py

Line 38 in 1a2f406

if aggregation:

To Reproduce

The following custom aggregation should output the tensor of values without aggregation but returns 1, the max of the values.

>>> from torchmetrics.retrieval import RetrievalNormalizedDCG
>>> indexes = tensor([0, 0, 0, 1, 1, 1, 1])
>>> preds = tensor([0.2, 0.3, 0.5, 0.1, 0.3, 0.5, 0.2])
>>> target = tensor([False, False, True, False, True, False, True])
>>> ndcg = RetrievalNormalizedDCG(aggregation=lambda x, dim: x)
>>> print(ndcg(preds, target, indexes=indexes))
tensor(1.)
>>> ndcg = RetrievalNormalizedDCG(aggregation="max")
>>> print(ndcg(preds, target, indexes=indexes))
tensor(1.)
>>> ndcg = RetrievalNormalizedDCG()
>>> print(ndcg(preds, target, indexes=indexes))
tensor(0.8467)

[github-actions]

Hi! thanks for your contribution!, great first issue!