Add Classification SFT recipe

[yi-fireworks]

Add Classification SFT recipe
This PR introduces a new recipe in learn/finetuning/classification-sft that demonstrates a complete workflow for Supervised Fine-Tuning (SFT) on Fireworks AI.

Minimal Steps/Scripts:

• End-to-end classification task: Classifying support tickets into billing/hardware/software.
• Data Generation: Script to generate synthetic training (90%) and validation (10%) datasets.
• Robust Training Script: Handles dataset uploads, training configuration, and auto-cleanup of previous runs.
• Validation Integration: Configures the SFT job to use a validation set for tracking loss metrics.
• Deployment Best Practices: Includes updated deployment commands with min-replica-count 1 for immediate warm start.
• Post-Training Evaluation: Includes calc_metrics.py to calculate one-vs-rest Precision, Recall, and F1 scores on the validation set.
• Cleanup Utility: Provides a robust cleanup.py script to tear down datasets, models, and deployments.

The recipe is structured with a centralized config.py for easy customization and includes a comprehensive README.md guiding users through the entire process. This serves as a "Quick Start" template for users looking to fine-tune classifiers on the platform.

[pgarbacki]

It would be better to use BuildSDK instead of calling the commands.

[pgarbacki]

cc: @dphuang2

[pgarbacki]

Is this the best way to do it or should we be using eval protocol? cc: @benjibc

[pgarbacki]

It may be better to put all this code into an ipython notebook that can be structured as an interactive tutorial.