Move parameter_init to train stepper configs

[jpdunc23]

Continuation of the separation of training-specific concerns from inference stepper configs, building on #809.

This PR makes backwards-incompatible changes affecting fme.coupled training and fine-tuning configs:

• Training (init from separate uncoupled component checkpoints): Each component's parameter_init: ParameterInitializationConfig is now configured using its respective ComponentTrainingConfig on the CoupledTrainStepperConfig.
• Fine-tuning (init from a coupled stepper checkpoint): Moved parameter_init: CoupledParameterInitConfig out of CoupledStepperConfig and into CoupledTrainStepperConfig.

Existing fme.ace training YAML configs will continue to work without changes, with parameter_init now transferred to TrainStepperConfig via StepperConfig.get_train_stepper_config(). This backwards compatibility will be removed in a future PR.

Changes:

• CoupledTrainStepperConfig now owns both CoupledParameterInitConfig and the per-component ParameterInitializationConfig (via ComponentTrainingConfig).

• StepperConfig.get_stepper() signature changed to accept an optional ParameterInitializer instead of a boolean flag

• TrainStepperConfig now owns parameter_init and builds both the initializer and the underlying Stepper

• Tests added