enable PCA on data to exploit sparseness

[jan-glx]

This PR

• adds a slot parameter to RunPCA.Seurat and RunPCA.Assay
  • this allows to run RunPCA on raw (counts) or normalized data (data) without running ScaleData first
• adds scale and center parameters to simplify use of implicit scaling/centering in RunPCA
  • Performing PCA with implicit scaling on a sparse matrix allows for a significant speedup (~6x) over explicit scaling through ScaleData (converts to a dense matrix) (see comment below for details)
• is based on Fix weight.by.var for approx=FALSE #2330 - which should be merged first
• is not yet ready for merge because:
  1. Should center, scale and slot parameter values be saved in the dimReduce object? Where?
  2. To be more consistent with ScaleData scale and center might be better names do.scale and do.center, scale and scale. and center parameters values should the be used as default if supplied, to keep existing code using these (through ....) functional
  3. for approx=FALSE only, RunPCA has by default performed centering, while this is only a problem if the user used do.center=FALSE in ScaleData, it is inconsistent and should be changed
  4. Implementation could be simplified by switching from irlba to prcmp_irlba if simplify and and optimize prcomp_irlba bwlewis/irlba#52 gets merged
  5. ?