Implement MCP methods

[Zethson]

A reimplementation of https://github.com/livnatje/DIALOGUE could be awesome.

Then we could also talk to @davidsebfischer et al. to see how far their MCP method is.

[Zethson]

TODO list

• [x] Implement DIALOGUE 2
• [x] Add DIALOGUE 3
• [ ] Separate load and calculate_multifactor_PMD calculation -> have two separate function calls for these; load should store stuff in the AnnData object already
• [x] Rename the MCPs to something better and not just 0, 1, ...
• [ ] Rewrite the MCP concatenation to ensure that we're not returning an AnnData object but modifying inplace
• [ ] Write tests for the functions
• [ ] Solve TODOs
• [x] Add plotting functions from DIALOGUE R
• [ ] Think about factoring all DIALOGUE 1, DIALOGUE 2 and DIALOGUE 3 steps into their own Python files aka create a private _dialogue module
• [ ] remove new_mcp_scores from multilevel_modeling and expose calculate_cca_sig
• [ ] ensure calculate_cca_sig and multilevel_modeling can be called separately
• [ ] Store the most used parameters in the DIALOGUE object -> e.g. sample etc
• [ ] hierarchical modeling only returns results for 1 cell type -> why?!? Is the result being overwritten?
• [ ] results from hierarchical modeling should not return sig_genes per cell type pair but rather per cell type
• [ ] make sure user is aware that n_UMI needs to be logged if .X is logged
• [ ] multilevel_modeling does not allow for confounder to be None -> code breaks (ignore type hints)