Liana-py

Changes
- Fixed an issue where interactions with complexes will not filtered be according to
`expr_prop` for some methods. I now filter twice - once via `.filt_liana_pipe`
for computational speed, and once after `recomplexify` to also remove the
complexes with `expr_prop` <= X. Will now also filter `crosstalk_scores` to `expr_prop`.

- In the edgecase of complexes with subunits with equal expression, LIANA's internal
methods will not arbitrarily discard duplicate complex interactions.

- Will now return `expr_prop` for each method. Note that this information is
discarded by `liana_aggregate`.

- `liana_doplot` function is now more explicit in the way interactions are
selected. Will now take `topn` and return the highest ranked interactions.
Size of dots is also more distinguishable by default and the user can now
pass a customizable value for the size range.

- Added a `rank_method` helper function to rank single methods according to
`specificity` and magnitude.

- Removed ~20 bad quality interactions from the `Consensus` resource.

- Minor changes on filtering SCE object in `liana_pipe` to ensure all
complex subunits are present in the sce

New Implementations
- Untargeted between-condition (context/sample) decomposition of cell-cell
communication latent patterns /w `tensor_cell2cell`. Makes use of `basilisk` to
automatically set-up a conda env for liana.
- added `min_cells` parameter to `liana_wrap`, to exclude any cell identity
which does not pass a minimum cells threshold.

Changes
- Mouse Consensus resource is now provided by default.
- The intracellular OmniPath vignette was removed. An updated and more user-friendly one
will be provided in next updates. In the meantime, the old one can still be downloaded
from [drive](https://drive.google.com/file/d/1lqxHhmz0Jq7eIuQAe0SxvInGgo2U-RlC/view?usp=share_link)
- Source and Target titles are now plotted by the `liana_dotplot`
- added explicit error if `idents_col` was not found in metadata/colData

Changes
- Reduced dependencies (specifically `Seurat` and `OmniPathR`)

Minor Changes
- testthat tests external methods only if requested explicitly
- Readme updated - clarified and accordingly describes the `Consensus` resource as default


New Features
- `idents_col` is can now be explicitly passed to `liana_wrap`, if not provided defaults to the active
idents/colLabels for SCE and Seurat, respectively.
- `verbose` param allows to omit any messages and warnings from LIANA
- `assay` can now be passed explicitly when working with a Seurat object, defaults to the active one otherwise

Changes
- Re-implemented the `RRA` method from Kolde et al., 2012, as a consequence of
the removal of the `RobustRankAggregate` package from CRAN.

- Integrate `generate_homologs` with OmniPath's `homologene` database.
This allows homology conversion by simply passing an organism ID. Also, handles
complicated cases, such as complex subunits with one-to-many mapping homologs.

New Implementations
- Frequency Heatmap available via the `heat_freq` functions, added due to being common requests.
This heatmap was inspired by CellPhoneDB and CellChat.

Changes
- Extended basic tutorial to accommodate new heatmap plots.

Minor changes
- Allow labels to be passed to `liana_dotplot`
- Cleaned up docs, dependencies, examples, and warnings

Changes

- LIANA will now use the `Consensus` resource by default. This is a highly-literature supported resource, generated using similar
filtering steps as the 'OmniPath' (old default) resource. This resource is similar in size (~4,700 interactions), but contains a
higher complex and curation content.

- All resources might show some very minor changes related to an update of UniProt IDs and homology-conversion improvements.

- LIANA now uses `mean0` to account for heteromeric complexes, i.e. the mean is computed, unless there is a value of 0, then 0 is returned.
This means that any complex, the subunit of which is not expressed is filtered. LIANA now also appropriately accepts any custom function to
account for complexes.

- `liana_aggregate` now groups by ligand.complex and receptor.complex as well as the subunits,
and hence returns a all of those columns


Minor Changes

- Added option to show complexes on dotplot and is now the default option

- Documentation improvements

- `decomplexify` function is now exported

- `liana_aggregate` will no longer return a median_rank, it's largely redundant.

- re-arranged the column order of `liana_aggregate` due to the addition of .complex columns

- Replaced min0 (used to obtain closest to 0 value) to min -> relevant for z-scores used in Connectome.

Bugs

- Complexes with missing subunits are not correctly assigned as 'missing' and hence filtered/treated as non-expressed.

- Fixed a bug where LIANA will return the minimum subunit expression, instead of the mean for some methods.
This stemmed from not properly passing the incorrect `complex_policy` to certain methods, i.e. they were getting a hard-coded value instead.

- Remove `decomplexify` logical from `liana_call` and `liana_pipe` -> redundant.

- edge case fix: liana_aggregate should now rank interactions with the same subunits, but coming from different complexes seperately