docs: add hybrid device-apiserver design for PR #718 + #720 merge

Comprehensive design document capturing:
- In-memory cache architecture (replacing Kine/SQLite)
- Memory safety bounds sized for Vera Rubin Ultra NVL576 (1024 GPUs)
- Readiness gate blocking consumers until providers register
- Watch broadcaster with slow subscriber eviction
- What to take from each PR
- Implementation phases
- Authorship preservation requirements

This document serves as the spec for cherry-picking #720 onto merged #718.

Co-authored-by: Carlos Eduardo Arango Gutierrez <carangog@redhat.com>
Co-authored-by: Dan Stone <danstone@nvidia.com>

Author: ArangoGutierrez