[Feature] Implement TopK wrapper class for metrics

🚀 Feature

(Follow up to #3530 and #2631)

After getting in the HitRate@top_k metric for recommendation systems, it makes sense to have Recall@top_k as well. To implement this metric, as discussed we can add a top_k argument to existing Recall metric instead of creating a new class:

The idea is to use the _prepare_output method of _BasePrecision_Recall to convert our data to the required format that can then be used directly with same logic of if self._average == "samples": of Recall class:
for the true_positives, fetch the top-k indices of y_pred and gather those elements from y.
y can then also be directly be passed as actual_positives.
https://github.com/pytorch/ignite/blob/24535a0bb078398d1374657ce6f8c1f583cd467a/ignite/metrics/precision.py#L60

_, top_indices = torch.topk(y_pred, k=self._top_k, dim=1)
true_positives = torch.gather(y, 1, top_k_indices)

then in https://github.com/pytorch/ignite/blob/24535a0bb078398d1374657ce6f8c1f583cd467a/ignite/metrics/recall.py#L229
we can use this same logic for final metric calculation

Start by reading ignite/metrics/precision.py, especially _BasePrecision_Recall._prepare_output, and ignite/metrics/recall.py around the Recall calculation. Compare the requested top-k behavior with the existing Recall logic and the linked HitRate work; done means Recall supports the requested top-k calculation consistently with the stated approach.