Port summarizer for Lotka-Volterra from numpy to PyTorch

[alvorithm]

The modules lotka_volterra_pytorch and mg1_pytorch both import from a non-checked-in module, summarizers. By default they will try to summarize observations (https://github.com/mackelab/sbi/blob/7f62e3c23a1cc56ffd92f6355636e61934c80881/lfi/simulators/lotka_volterra_pytorch.py#L74, https://github.com/mackelab/sbi/blob/7f62e3c23a1cc56ffd92f6355636e61934c80881/lfi/simulators/mg1_pytorch.py#L23) and fail here.

These pytorch simulator varians do not seem to be in use at the moment.

[janfb]

Looking at the numpy implementation in lotka_volterra.py we see that there is a Stats() class acting as a summarizer.

I think Conor first made it possible to simulate with and without summary stats, therefore the argument summarize_observations=True, and the case

https://github.com/mackelab/sbi/blob/39a0dbf792f7864383290d8ead69b1a7bfe5cda5/sbi/simulators/lotka_volterra_pytorch.py#L73-L74

and then got rid of this feature to calculate summary stats by default.

If we need the pytorch implementation then we would have to write the Stats() class in PyTorch.

However, in all current examples and tests the numpy version is used.

[jan-matthis]

Would be good to port the summaries over to PyTorch, imo

[alvorithm]

If Lotka-Volterra is going to disappear from sbi, then we should close this? Alternatively, this issue will be dealt with in different form when adopting pytorch-lightning

[jan-matthis]

If Lotka-Volterra is going to disappear from sbi, then we should close this?

Yes, think this can be closed