MinMetric produces incorrect result when used with MetricCollection and compute_groups=True

[tshead2]

🐛 Bug

Using MinMetric and MaxMetric in a MetricCollection with compute_groups=True, the minimum metric produces an incorrect result (it returns the maximum).

To Reproduce

Steps to reproduce the behavior...

>>> import torchmetrics
>>> m = torchmetrics.MetricCollection({
...   "training/min": torchmetrics.MinMetric(),
...   "training/max": torchmetrics.MaxMetric(),
... }, compute_groups=True)
>>> m.update(1)
>>> m.update(2)
>>> m.compute()
{'training/max': tensor(2.), 'training/min': tensor(2.)}

Expected behavior

The output from this example should be

{'training/max': tensor(2.), 'training/min': tensor(1.)}

... setting compute_groups=False produces the correct result.

Environment

• TorchMetrics version 1.0.0, installed with pip
• Python 3.10.8, PyTorch 2.0.0, installed with conda
• Linux OS

[github-actions]

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