fix: optimize dockerfile to reduce image size

[TimPietrusky]

Motivation

• Implemented multi-stage build pattern to reduce the final image size
• Added selective package copying from builder stage to minimize redundant dependencies
• Optimized CUDA library handling by keeping only essential libraries
• Improved cleanup of cache files and temporary data
• Implemented better environment variable settings for reduced disk usage

Known Limitations

• The 'no space left on device' error may still persist in environments with limited disk space
• Further optimization might be needed depending on the specific deployment environment

Technical Details

• Uses multi-stage build with axolotlai/axolotl-cloud:main-latest as base
• Selectively copies only required Python packages from builder stage
• Implements thorough cleanup of unnecessary CUDA libraries while preserving essential ones
• Optimizes package installation with no-cache settings
• Maintains all required functionality while reducing image footprint