Minimal LoRA implementation

[jlamypoirier]

✨ Description

Same as #180, but without the infrastructure changes needed to support LoRA/PEFT for MLP, TP, etc. Part of #149

• Create Lora wrapper and Lora (peft) config
• Support Lora for query and key_value layers, and add option to selecting which of them to apply lora on.
• Add option to freeze non-lora weights
• Add support for freezing entire layers/stages (disable backward pass if it's the first one(s), missing in Support frozen weights #185)
• Make it work on a manual test
• Prevent lora wrappers from messing up parameter names (breaks checkpoint conversion)
• Support Lora on key and/or value separately
• Checkpoint conversion (separate PR)
• Add test(s)

🔍 Type of change

Select all that apply:

• 🐛 Bug fix (non-breaking change that addresses a specific issue)
• 🚀 New feature (non-breaking change that adds functionality)
• ⚠️ Breaking change (a change that could affect existing functionality)
• 📈 Performance improvement/optimization (improves speed, memory usage, or efficiency)
• 🛠️ Code refactor (non-functional changes that improve code readability, structure, etc.)
• 📦 Dependency bump (updates dependencies, including Dockerfile or package changes)
• 📝 Documentation change (updates documentation, including new content or typo fixes)
• 🔧 Infrastructure/Build change (affects build process, CI/CD, or dependencies)

[tscholak]

Thanks, this is great!
import-export of HF-compatible PEFT weights is indeed better to keep for a follow-up ticket. thanks

[jlamypoirier]

@tscholak Opened #204. Will merge this one once #185 is merged to main.