codex tok/s double-counts reasoning tokens in generated-token throughput #1079

Split from #1075/#1078: the throughput path has the same reasoning-inclusion bug as the cost path fixed in #1078 — activeGeneratedTokens/taskGeneratedTokens in src/providers/codex.ts and generatedTokens in src/codex-throughput.ts:403 add reasoning tokens to output that already contains them, so displayed tok/s reads high (~+34% on a corpus where reasoning is ~1/3 of output). Display-only, no cost impact. Fix means rewriting ~7 existing throughput assertions; use billableOutputTokens('codex', …) from #1078 so the definition can't drift from the cost path.