Add QLoRA and advanced PEFT presets

[SahilKumar75]

Problem

Many users will need low-memory fine-tuning paths for larger open models.

Scope

• Add QLoRA presets with quantization-aware configuration.
• Add presets for common hardware profiles.
• Support future PEFT techniques through a shared configuration model.

Acceptance Criteria

• Users can select QLoRA as a technique.
• The UI explains the tradeoff between memory, speed, and quality.
• Presets populate safe default values that users can adjust.