fmwork Input Options Enhancement

[WarningRan]

Current Status

• Primarily rely on synthetic data for experiments
• Limited support for real datasets (mainly ShareGPT)
• No batch-level correlation control
• Insufficient diversity in test scenarios

Planned Improvements

• Add support for multiple real-world datasets beyond ShareGPT
• Implement batch-level correlation between samples in real dataset data

[WarningRan]

have initial implementation batch-level correlation between samples in real dataset data in close repo.