Fix GKD Liger teacher routing

[taking-lying-flat]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Fix the teacher routing and validation logic for Liger GKD loss.

Previously, _prepare_liger_loss() was called before _setup_teacher(). As a result, Liger GKD loss was initialized before the trainer resolved whether GKD was using a local teacher model, a teacher API, or self-distillation.

This PR:

• Moves Liger loss initialization after teacher setup, ensuring that it uses the resolved teacher configuration.
• Requires a local teacher model when use_liger_kernel=True.
• Raises a clear error when Liger GKD loss is used with teacher API or self-distillation, as these modes are not supported.
• Raises a clear error when OPSD samples are used with Liger GKD loss.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[hjh0119]

plz pass the lint test