fix(codex): support exec optimizer backend

[TheGameVIX]

Summary

• allow codex_exec as an optimizer backend and make --backend codex route both optimizer and target through Codex exec
• route optimizer calls through skillopt.model.codex_backend, including token tracking and Codex exec config
• send SkillOpt-Sleep Codex prompts through stdin (codex exec -- -) so multiline replay prompts work reliably on Windows
• update backend docs/config references and add regression tests

Why

This is a stacked follow-up for microsoft#101. That PR adds Windows/Codex plugin support, but the normal SkillOpt train/eval path still left the optimizer on openai_chat when users selected --backend codex. That meant users could still hit API-backed optimizer calls instead of Codex exec.

During Windows smoke testing, skillopt_sleep --backend codex also exposed a Codex CLI prompt transport issue: multiline prompts passed as argv could produce bad replay responses. Passing the prompt through stdin keeps the same codex exec backend while avoiding Windows argv parsing edge cases.

Validation

• python -m pytest -q (170 passed, 5 skipped)
• python -m mkdocs build --strict
• python -c "import sys; sys.argv=['train.py','--config','configs/searchqa/default.yaml','--backend','codex']; from scripts.train import parse_args, load_config; cfg=load_config(parse_args()); print(cfg['optimizer_backend'], cfg['target_backend'])" prints codex_exec codex_exec
• Windows Codex smoke: python -m skillopt_sleep dry-run --project "D:\Documents\AI\manhwa_recap_local_ui\manhwa_recap_local_ui" --backend codex --tasks-file "<reviewed smoke tasks>" --progress --json exited 0 with reviewed tasks and no adopt