fix: filter full recipe template from serverless train( )

[mujtaba1747]

Title: fix: filter full recipe template from serverless train()

Summary

• Serverless CreateTrainingJob limits HyperParameters to 100 members, but the fully resolved
  recipe template can contain hundreds of internal leaf keys the user never touched, causing API
  validation failures.
• Adds only_user_overrides flag to _apply_recipe_to_hyperparameters() so serverless trainers
  only forward keys the user explicitly provided (direct assignments, overrides dict, or recipe
  YAML).
• Applies the filter in SFTTrainer, DPOTrainer, and MultiTurnRLTrainer. RLVRTrainer retains
  full recipe passthrough since it renders into a YAML file with no key-count limit.

Test plan

• Integ and unit tests to validate new logic
• Verify serverless training jobs (SFT, DPO, MultiTurnRL) no longer fail with "too many
  hyperparameters" when using a large recipe template
• Confirm user-provided overrides are still forwarded correctly
• Validate RLVRTrainer still passes the full resolved recipe (no regression for serverful
  paths)

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.