Changing the use_raw behavior without breaking things

It is common to store raw counts (=unnormalized) of all measured genes under adata.raw, while having normalized and unnormalized expression of a subset of genes (might be only protein coding genes, or all genes except ribosomal and mitochondrial etc) at adata.X and adata.layers['counts'] respectively. This however gives rise to a lot of trouble in plotting since visualizing raw counts is not a great idea due to the dynamic range. It is super annoying to pass use_raw=False to a lot of functions. Furthermore, weird rank_genes_groups outputs as a result of raw counts might go unnoticed because of this. (happened to me many times)

I think there was a discussion somewhere about switching to use_raw=False by default in all functions, but this may potentially break things since it's a significant behavior change. This might be reasonable for Scanpy 2.0, but not in 1.x I assume.

My suggestion is to have a use_raw option under sc.settings (i.e. the global ScanpyConfig instance) which is None by default, and can be set to False (e.g. sc.settings.use_raw=False) which would then affect all the functions with use_raw argument. This way we don't break the behavior but still have a reasonable way to turn this thing off :)

Let me know what you think.

Review the global ScanpyConfig instance exposed as sc.settings and the functions that accept a use_raw argument. Determine how a None default should preserve existing behavior while sc.settings.use_raw=False affects those functions, then verify plotting and rank_genes_groups behavior. The issue names no specific files or tests.