feat(report): curate the multisample PCA feature set and add a QC overview

[CarlosBlancoGo]

Summary

The multisample PCA ran prcomp() on every numeric column of the cell summary
(348 columns). With scale.=TRUE every column carries equal weight, so a
property spread across many correlated columns (the junction breakdown,
structural categories) dominates the components over one held in a single column,
and PC1 tracks the largest column family rather than the differences between
samples.

This PR runs the PCA on a curated feature registry, adds a per-sample QC overview
heatmap of the medians, and adds a --pca_features flag to override the set.

Details

Six blocks, one interpretable feature family each:

Block	Features
Yield & detection	reads/transcripts per cell, annotated genes per read, UJCs per read, mitochondrial %
Transcript length	per-cell median length
Structural categories	FSM, ISM, NIC, NNC, genic genomic, antisense, fusion, intergenic, genic intron (proportions)
Splice junction composition	known/novel × canonical/non-canonical junction proportions
Good-quality indicators	TSS annotation / PolyA motif / CAGE peak / TSS-ratio / short-read SJ support
Bad-quality indicators	RT-switching, intrapriming, reads with non-canonical SJ, predicted NMD

The good/bad blocks hold the same features the per-sample report's "Good/Bad
quality features" figures declare, so one taxonomy spans both reports.

Yield's gene and UJC features are depth-normalised per cell — divided by the
cell's own read (or transcript) count before the median is taken — so they
measure detection efficiency rather than sequencing depth, which is kept as its
own feature. Flag-gated features (ORF, polyA, CAGE, coverage) enter the registry
only when they vary within every sample, since cell_metrics writes them with a
constant sentinel when the flag was unused.

In the QC overview heatmap, each feature (row) is coloured by a robust z-score across samples,
(x - median) / (1.4826 * MAD), with a mean/sd fallback when MAD is zero and a
scale limit of min(4, max|z|) so an ordinary cohort still fills the gradient
while one outlier cannot wash the rest out.

It is interactive in the HTML report (ggiraph, per-tile tooltips giving the
sample, feature, median and z-score) and static in the PDF. Layout tracks cohort
size — fonts, label angle, canvas, and per-tile numbers dropped past ~8 samples —
columns face by color_group when the design supplies one and cluster otherwise,
and block-name strips wrap so each word stacks on its own line. ggiraph is used
rather than plotly because it renders the real ggplot to SVG and preserves
facet_grid(space = "free"), keeping the block heights proportional to their row
counts. r-ggiraph is added to the environment; without it the report falls back
to the static figure.

--pca_features <file> (pipeline and generate_multisample_report.py) takes one
column name per line and replaces the curated set.