fix: Fix device mismatch when DPO runs validation at start with CPU offload (Nemotron MoE)

[RayenTian]

What does this PR do ?

Summary
When DPO is run with val_at_start=true and policy.dtensor_cfg.cpu_offload=true, validation can crash on MoE models (e.g. NVIDIA-Nemotron-3-Nano-30B-A3B-BF16) with:

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

The failure occurs in the model’s MoE gate (e.g. get_topk_indices: scores on CUDA, e_score_correction_bias on CPU).
Dense models (e.g. Qwen2.5-7B) do not hit this under the same settings.
This PR fixes the bug by calling policy.prepare_for_lp_inference() before running the initial validation when val_at_start is enabled, so that all buffers (including MoE gate buffers) are on CUDA before any reference-policy logprob computation.

Issues

Related to #1922

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

• ...