Fix MoE multi-lora training

[tastelikefeet]

PR type

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

PR information

Write the detail information belongs to this PR.

Experiment results

Paste your experiment result here(if needed).

[gemini-code-assist]

Code Review

This pull request introduces support for configurable target_modules in multi-LoRA implementations for both Megatron and Transformers, defaulting to 'all-linear'. It also adds 'Qwen3.6' to the model template mapping and implements logic to preserve the 'all-linear' configuration during model saving. Feedback focuses on ensuring state consistency during the save process by using a try...finally block and restoring performance-critical checks in activate_adapter to avoid redundant calls to set_adapter.