Add network, sankey, COG bar chart, and ego-network visualizations

[eboyer221]

Summary

This PR adds several new visualizations to the dashboard, all integrated with the existing data sources.

New visualizations

• Drug-feature network (new "Network" tab) — interactive networkD3 force-directed graph linking drugs/drug classes to their top ML-predicted features. Edge thickness encodes feature importance. Optional toggles add cluster and COG tiers from the annotations parquet.

• Resistance flow Sankey (Metadata tab) — multi-tier diagram showing phenotype → drug class → drug → country → host → isolation source. Flow widths are record counts. A drug-class multi-select lets the user focus the diagram (defaults to top 3 classes by record count).

• Cross-model accuracy distributions (Model holdouts tab) — box-with-jitter plots of balanced accuracy by drug class, split by country and time, colored Same vs Different (self-eval vs cross-tested). Added as a new "Accuracy distributions" sub-tab.

• Cross-model performance heatmaps (Model holdouts tab) — country and time heatmaps now render side by side in a unified "Model performance" sub-tab, using a purple gradient (#f2f0f7 → #54278f) with axes labelled "Train data" / "Test data". Titled "Cross-country performance" and "Cross-time performance".

• COG bar chart (Feature Importance tab) — horizontal bar chart of the most common COGs across the currently displayed features, on both "Across bugs" and "Across drugs" sub-tabs.

• Feature ego-network (Feature Importance tab) — small networkD3 graph that updates when a row is clicked in the feature importance table, showing the selected feature → its cluster (BV-BRC link) → its COGs (NCBI links).

Annotation enrichment

• New load_feature_annotations() and enrich_with_annotations() helpers read cluster_feature_COG.parquet and join feature ids on Variable → feature (splitting on _ to extract Pfam ids), collapsing multiple COGs per feature into comma-separated cells.
• New load_feature_name_map() reads gene_names.parquet / domain_names.parquet / protein_names.parquet from amrdata_root (defaults to ~/amRdata/data if it exists), giving heatmap rows like group_6367 (Glutamate decarboxylase (EC 4.1.1.15)) instead of opaque IDs.
• Feature importance table now shows cluster (linked to BV-BRC) and COG (linked to NCBI COG) columns. The unhelpful COG_name hyperlink was removed.

Other changes

• launchAMRDashboard() gains an optional amrdata_root argument for the name-map / annotation lookup.
• networkD3 moved from Suggests to Imports.
• Included cluster_feature_COG.parquet for Shigella flexneri in inst/extdata/ so the new visualizations work out-of-the-box with the demo data.

Test plan

1. From the package root, open R and run:

   devtools::load_all(".")
   app <- launchAMRDashboard()
   shiny::runApp(app)

[eboyer221]

Closing in favor of #20, which includes all these changes plus the package rename from amRshiny to amRviz.