gex_only=True defaults in sc.read_10x_h5

This might be a bit of a rant, and I'm aware there are some good arguments for the way things are... but I just wasted 4 hours of my life because I wasn't aware of the default gex_only=True in sc.read_10x_h5().

Just wanted to flag that if scanpy support for multimodality becomes a thing, then this default may need to change to prevent frustration. I think moving read/write functionality to AnnData was already discussed at some point. For multimodal support, this might become more important again. If this is moved, then read_10x_h5 should probably not default to gex_only=True anymore.

Would it already be worth either making gex_only a required input? For backward compatability it could also just trigger a logging warning for now. I do think that this is quite an important thing to let people know with more and more 10X Multiome data being generated now (and CITE-seq for that matter).

Start at the sc.read_10x_h5() entry point and inspect how the gex_only=True default affects 10X multimodal data. Review the proposed relationship between read/write functionality and AnnData, then determine a backward-compatible behavior or warning. Done means the chosen behavior is implemented and clearly communicates how multimodal inputs are handled.