Initiate the deconvolution notebook draft

[sjspielman]

Towards #980

This PR begins the draft of the deconvolution notebook. There's kind of a lot here already, which was somewhat because I was on a roll and also somewhat intentional because I wanted to hear opinions on scope and I thought that would be easier to do seeing more laid out. It could probably use some more headers, but I am waiting for a first round of review before I add more structure since the content might shift. Here's what the notebook includes:

• plot some marker genes to set expectations (this can be pared down further too, let me know)
• peek at the reference dataset
• set up to run RCTD
  • several TODOs in this section as I was debating about the right level of detail for both text and for looking at the output objects. Some of the content is here just for sharing during review, to get some feedback on how much detail to present
• run in doublet mode, plot
  • includes one chunk plotting Tumor expression so they can type out the function, and then a populated chunk to plot all cell types. we type it for them for the rest of the notebook
• run in full mode, plot
• explore the full mode results, specifically:
  • look at the lymphoid proportions a bit more in depth to conclude it may be noise
  • calculate the total number of cell types per spot to get a birdseye view of heterogeneity
• exploration of "what happens if we use dominant cell types" - make a barplot of tumor-dominant to show that we may lose information by collapsing to a dominant

Other items beyond this that were in the outline (and actually slated to potentially get used!) include: comparing to clusters & co-occurrence. I was thinking to get a cluster comparison into the instruction notebook, and leave co-occurrence for an exercise (I would expect to fill in a lot of chunks for them). But again, I first wanted to file this PR to hear some feedback about current scope before I start making such decisions. For example, I am also flirting with move the dominant cell type section into an exercise because I think it would suit that nicely, and we can put a cluster comparison in its place in the instruction while keeping roughly this overall length. But, on the other hand, is this point important enough to want to keep it in instruction? Maybe! In any case, all kinds of thoughts welcome!

Notebook is here: 03-deconvolution.nb.html