Add multi-GPU precompute test coverage for KTO

#6403 fixes precompute for DPO **and** KTO but only adds a DPO distributed test (`test_dpo_precompute_ref_log_probs`, ddp/zero2/zero3). Add the KTO mirror so the duplicated paths can't silently diverge: `trl/scripts/kto.py --precompute_ref_log_probs --eval_strategy epoch` (two-pass, the case that broke ZeRO-3), same configs. Check `kto.py` wires `eval_dataset` like `dpo.py`; exclude FSDP2 until Issue 1 lands.

Start with the existing test_dpo_precompute_ref_log_probs distributed coverage and compare its ddp, zero2, and zero3 configurations. Then inspect trl/scripts/kto.py alongside dpo.py for eval_dataset wiring, and run the KTO equivalent with --precompute_ref_log_probs --eval_strategy epoch. Done means the KTO two-pass case is covered for the same configurations while FSDP2 remains excluded.

Written by the indexing model from the issue text.