Add CUDA wishlist document for persistent GPU actor features

Comprehensive analysis of CUDA features that would enable true
persistent GPU actors, based on RingKernel implementation experience.

Key proposals include:
- Native host-kernel signaling (replace polling)
- Kernel-to-kernel mailboxes (first-class messaging)
- Dynamic block scheduling (work stealing)
- Persistent kernel preemption (cooperative)
- Checkpointing and migration (fault tolerance)
- Extended cooperative groups (hierarchical sync)
- Memory model enhancements (SC atomics)
- Multi-GPU persistent kernels

https://claude.ai/code/session_01TD1CHULcRkSAJ1KUqyhpF9

Author: claude