feat: implement Small Model Mode (Phases 1 & 2) #391

[kaustubha07]

This PR implements the first two phases of the "Small Model Mode" initiative (#391), focusing on optimizing Nanocoder for local, low-parameter models.

Key Improvements:

• Slim System Prompt: A dedicated ~300-word prompt for high performance on small models.
• Tool Subsetting: Predefined profiles (minimal, code-edit, etc.) to reduce model confusion.
• Schema Simplification: Automatic recursive stripping of descriptions/examples from tool schemas.
• Smart Retry: Intelligent fallback to Small Model Mode and XML parsing on tool errors.
• Auto-Detection: Regex-based activation for models like Llama 3B, Gemma 2, Phi, and Qwen.

Verification:

• Targeted unit tests for detection and simplification logic are passing.
• 100% compliance with project linting, formatting (Biome), and type checks.
• Verified on Windows; pre-existing environment-related AVA failures were documented but no new regressions were introduced.