Add a regression test for the non-FSDP ORPO path

[kashif]

Nice fix! The only thing missing from linkedin#1229 / linkedin#1230 was a test that locks it in, so here's one to go on top of your branch.

It builds a tiny Llama and calls concatenated_forward on a plain (non-FSDP) model. Before your change this path went unconditionally through _FSDPForwardRedirection, which asserts the module is FSDP and blew up on a single GPU — exactly the crash from linkedin#1229. The test just checks it runs and returns a finite loss.

I kept it narrow on purpose: it stubs concatenated_inputs and builds the trainer with __new__, so it only exercises the forward path you touched and doesn't drag in the whole dataset/tokenizer pipeline. Guarded with importorskip("trl") since the trainer subclasses trl's ORPOTrainer, and it uses infer_device() so CI runs it on GPU.

Feel free to squash it into your branch.

[kashif]

Folded the trl-compat work into this branch so everything ORPO lands together (closed linkedin#1244 in favor of this):

• orpo_trainer.py: import ORPOTrainer from trl.experimental.orpo with a fallback to the old trl.trainer path — recent trl moved ORPO into experimental, so the old import fails outright otherwise.
• run_orpo.py + docs/Examples.md: same fallback for ORPOConfig (the docs snippet was flagged by Copilot on Import ORPO from trl.experimental linkedin/Liger-Kernel#1244).
• Added the skipped regression test for the non-FSDP path (importorskip("trl"), so it's a no-op in CI where trl isn't installed, and runs where it is — verified passing locally on trl 1.5 + GPU).

No new deps: trl stays optional, nothing added to setup.py. ruff clean.