Revert xfail for MoE LoRA tests now that peft#3602 fixed the autocast dtype bug

[albertvillanova]

This PR reverts the temporary xfail markers added for the MoE LoRA tests, now that the underlying bug has been fixed upstream:

• Fix LoRA target_parameters raising a dtype error under autocast peft#3602

Fix #6914.

Motivation

peft's LoRA parametrization for MoE expert parameters returned bf16 under autocast, which register_parametrization rejects, so the MoE LoRA tests were marked xfail as a temporary hotfix (#6915) to keep CI green. The bug has since been fixed upstream in peft#3602.

Solution

Revert the hotfix commit that added the xfail markers, restoring the tests to their normal (non-xfail) state.

Changes

• Remove the xfail markers from the MoE LoRA tests in test_dpo_trainer.py, test_grpo_trainer.py, test_kto_trainer.py, test_reward_trainer.py, test_rloo_trainer.py and test_sft_trainer.py
• Remove the now-unused Version and peft imports from test_reward_trainer.py

---

Note

Low Risk
Test-only change with no production code paths; CI may fail if the environment still uses a peft version without the upstream fix.

Overview
Reverts the temporary xfail markers on MoE + PEFT/LoRA training tests now that peft#3602 fixes the autocast dtype issue that broke register_parametrization (closes #6914).

Affected suites: DPO, GRPO, KTO, Reward, RLOO, and SFT (test_train_moe_with_peft_config / test_train_moe_peft_model where applicable). Those tests run as normal failures/pass again instead of expected failures on peft dev releases.

In test_reward_trainer.py, drops unused packaging.version.Version and top-level peft imports that only served the xfail condition.

^{Reviewed by Cursor Bugbot for commit dec5654. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[albertvillanova]

Merging this PR as a hotfix: https://github.com/huggingface/trl/actions/runs/32867925767/job/97867656964

FAILED tests/test_dpo_trainer.py::TestDPOTrainer::test_train_moe_with_peft_config - [XPASS(strict)] peft's LoRA parametrization for MoE expert parameters returns bf16 under autocast, which register_parametrization rejects (see #6914).
...