OA-only mode: remove dragen_hrd dependency, keep it optional not deleted (sash #56)

[qclayssen]

Summary

Supports sash's OA-only mode (driven by the Grimmond Lab SBX evaluation — SBX is Roche's sequencer, samples arrive with no DRAGEN output) by making DRAGEN HRD scoring genuinely optional in the cancer report, rather than removed outright.

The first commit on this branch (d012f30) removed dragen_hrd from the report entirely. Since cancer_report.Rmd is the single shared template behind gpgr's canrep CLI — used for every report gpgr generates, not just OA-only ones — that would have silently dropped DRAGEN HRD scoring from regular, DRAGEN-capable sash runs too. It also left the report rendering a "DRAGEN" section always filled with NA instead of omitting it.

This commit fixes both:

• Restores dragen_hrd as an optional parameter end-to-end (cancer_rmd(), --dragen_hrd CLI arg, render.R) — it was already optional pre-removal, so this is a revert, not new API surface.
• hrd_results_tabs() now takes dragen_res = NULL and genuinely omits the DRAGEN column/spanner when absent, instead of showing it blank. Also fixes a stray left-border artifact that appeared on the CHORD column once it became the leftmost group (caught during review — the function had only ever been exercised with DRAGEN present).
• The DRAGEN description block, HRD summary row, and the hrd/*-dragen output artifact are all now conditional on whether DRAGEN was actually supplied.
• pcgr_format_categories() (PCGR skip/selection status parsing, sash umccrise canrep: flag if sample is hypermutated #52): fixed a no-separator string join that could fuse adjacent categories together if bolt emits pcgr_selection_filters as a multi-element vector, added defensive validation so a malformed category string is skipped instead of aborting the whole render, and collapsed nested ifelse() chains to case_when+switch so the value/details strings can't drift out of sync.
• Added test coverage for hrd_results_tabs() (both branches) and pcgr_format_categories() — both were previously untested.

Depends on: sash #58, bolt #34 (companion PRs for the OA-only path itself — this PR only touches the cancer report layer).

Known, deliberate gaps (flagged for follow-up, not blocking)

• The cli_warn previously fired on any missing --dragen_hrd is not restored — it would now be spurious on every legitimate OA-only run, and there's no way to distinguish "intentional OA-only" from "accidental omission" at this CLI layer.
• hrd_results_tabs()'s dragen_res = NULL couples a general-purpose exported function to one caller's data-availability needs; a cleaner design would push that decision into the calling Rmd. Left as-is since re-architecting is bigger scope than this regression fix.
• dplyr::select() on a malformed (but non-NULL) dragen_res still throws an opaque tidyselect error — pre-existing behavior, confirmed unchanged by this diff.

Test plan

• Full test suite: 51 passed, 0 failed (devtools::test())
• New tests: test-hrd-results-tabs.R (DRAGEN present/absent, both render to HTML without error), test-pcgr-status.R (category parsing, malformed-input handling, multi-element vector join)
• Manually verified hrd_results_tabs() output shape and gt::as_raw_html() rendering for both branches
• Reviewed via 2x 8-angle multi-agent code review (medium effort): correctness, removed-behavior, cross-file impact, reuse, simplification, efficiency, altitude, conventions — all confirmed findings fixed