Create a `plotMarkerHeatmap` function

[LTLA]

... that focuses on the most useful markers for a particular test dataset. Specifically, we take all of the reference-derived markers and look at which of them are also DE in the test dataset and then we use those for visualizing expression.

Something like this:

for.viz # assume this is a test matrix
colnames(for.viz) <- seq_len(ncol(for.viz))
col.order <- order(predictions$labels)

all.markers <- metadata(predictions)$de.genes
uniq.lab <- unique(predictions$labels)
for (lab in uniq.lab) {
    # Fetching all markers for this label.
    keep <- rownames(test) %in% unlist(all.markers[[lab]])
    current <- for.viz[keep,,drop=FALSE]

    # Prioritize the markers with interesting variation in the test data for
    # visualization. If we only have one label, we use the most abundant markers.
    if (length(uniq.lab) > 1L) {
        interesting <- scrapper::scoreMarkers(
            current,
            predictions$labels,
            compute.auc=FALSE,
            num.threads = BiocParallel::bpnworkers(BPPARAM)
        )
        cohen.stats <- interesting$cohens.d[[lab]]
        o <- order(cohen.stats$min.rank)
        to.show <- rownames(cohen.stats)[o]
    } else {
        abundance <- Matrix::rowMeans(current)
        to.show <- names(abundance)[order(abundance, decreasing=TRUE)]
    }

    cat("\n\n##", lab, "{-}\n")
    pheatmap::pheatmap(
        current[head(to.show,20),col.order,drop=FALSE],
        breaks=seq(0, max(for.viz) * 1.01, length.out=26),
        color=viridis::viridis(25),
        annotation_col=data.frame(
            labels=predictions$labels, 
            row.names=colnames(current)
        ),
        cluster_col=FALSE
    )
}

The aim is to generate some nicer looking diagnostic plots without relying on users to manually configure pheatmap::pheatmap (or scater::plotHeatmap, for that matter).

[LTLA]

Done in 1520a39c59d792a2ab20ecc0b4febb700b3fd7e6.