Add GPU-accelerated raytracer (45x speedup)

[steveant]

Summary

Adds GPU acceleration using Numba CUDA, achieving 45x speedup over CPU multiprocessing (530x over single-core).

Performance

Tested on NVIDIA GB10 (Blackwell, Compute Capability 12.1):

Scene	Resolution	CPU (20 cores)	GPU	Speedup
twoballs	960x540	0.82s	0.25s	3x
manyballs	1920x1080	14.3s	0.32s	45x

Changes

• gpu_engine.py - Numba CUDA kernels with structure-of-arrays memory layout
• main_gpu.py - GPU entry point
• pyproject.toml - Modern Python packaging with dependencies
• examples/manyballs.py - Larger benchmark scene
• README.md - GPU documentation

Key optimizations

1. Binary PPM output (P6) - Original text I/O was 96% of runtime
2. Iterative ray tracing - Replaced recursion for CUDA compatibility
3. fastmath compilation - Faster floating-point operations
4. Coalesced memory access - Structure-of-arrays layout

The clean architecture of the original made GPU porting straightforward.