A little suggestion of update to make ROGUE suitable for large dgCMatrix data

[mimi3421]

Dear author,

I have used ROGUE in place of the default HVG algorithm for feature sellection in the fastMNN pipeline and found that it returned a more concise UMAP graph and a more reasonable result in functional analysis after clustering.

I make a little change to reduced the RAM footprint and summit here if someone need it.

As log(0+1)=0, I make a little change in the Entropy function so that only the non-zero log result will be updated when dgCMatrix is supplied as the input data matrix, which would greatly reduce the RAM consumption.