feat: add stratified whole-database signal scanning

[MadDERt]

Summary

• Add faers_phv_scan_stratified(), a stratified variant of the whole-database
  scanner: it splits the database into strata defined by categorical demo
  columns (sex, age_grp, occp_cod, ...) and builds one 2x2 contingency
  table per stratum x drug x event combination against the stratum-internal
  background (n_stratum).

Details

• Missing strata values are grouped into an explicit "Missing" stratum by
  default (.na_stratum = "keep"), or excluded with "drop"; a warning is
  issued when a stratification column is missing for >30% of reports.
• .min_a applies per stratum; strata columns must be categorical
  (character/VARCHAR) demo columns so both backends produce identical output.
• Memory backend: stratum key carried through the integer-coded, chunked join
  (reports are atomic per stratum, so chunk counts stay exact).
• duckdb backend: the whole stratified aggregation is pushed into a single SQL
  query with a strata CTE (CASE ... 'Missing' coalescing / WHERE for
  drop mode).
• No changes to existing scan.R / scan-db.R code paths.

Tests

• New tests/testthat/test_scan-strata.R: 94 assertions covering per-stratum
  2x2 correctness vs direct computation, keep/drop semantics, cross-strata,
  per-stratum .min_a, chunked vs unchunked identity, and memory-vs-duckdb
  identical() parity (pt / soc_name / hlgt_name).
• Full suite: 455 passed, no regressions. R CMD check shows no new issues
  (env-only items unchanged).