Iog-scaling of counts data

[magnusrattray]

Hi,

In the synthetic benchmark experiments we noticed that the counts data are log-transformed (using sc.pp.log1p(adata)) before being rounded to the nearest integer and passed to the GPcounts method (and possibly same for other methods). GPcounts works better if passed the actual counts. The log transformation leads to an unusual counts distribution, e.g. for a non-spatially variable gene the data can have lower variance than a Poisson distribution (under-dispersed) which would never happen in real data. This may affect other packages that use counts. My student has a pull request fixing the issue.