Parallel batch calls optimization

• Parallel batch calls with diagnostics = FALSE now return compact worker
  summaries rather than row-level solver metadata. Final values and aggregate
  numerical-failure warnings are unchanged, while PSOCK result transfer is
  substantially reduced.
• Liljegren performance runners now time the public default
  diagnostics = FALSE; residual, fallback, and worker-parity validation is
  performed separately outside timed sections.
• Documented how to choose external foreach parallelism versus in-package
  workers, including the requirement to avoid nested worker pools.
• Updated documentation generation to roxygen2 8.0.0 and bound affected
  roxygen blocks to their exported public functions.