Add Callback API for Training Lifecycle Events

[kapil27]

Problem Statement:

Training Hub currently lacks a callback mechanism for extending training behavior at runtime. HuggingFace Transformers provides TrainerCallback that allows users to hook into training lifecycle events (epoch start/end, step end, evaluation, etc.). Training Hub should provide similar extensibility.

Without callbacks, users cannot inject custom logic during training without modifying Training Hub internals or relying on external file polling workarounds.

Desired API

A callback system similar to HuggingFace Transformers:

from training_hub.algorithms.sft import sft
from training_hub.callbacks import TrainingCallback

class MyCallback(TrainingCallback):
    def on_step_end(self, args, state, control, **kwargs):
        # Custom logic here
        pass

sft(..., callbacks=[MyCallback()])

Use Cases

1. Progress tracking and monitoring
2. Integration with experiment tracking platforms (MLflow, Weights & Biases)
3. Custom training behaviors (early stopping, dynamic learning rates)
4. Real-time updates to external systems

Reference

HuggingFace Transformers TrainerCallback: https://huggingface.co/docs/transformers/main_classes/callback