Add single-endpoint coordinated key-range sharding PoC

[haorenfsa]

What changes

A single Kubernetes endpoint can now split one resource collection across independent data shards. Opt-in coordinator + shards stores immutable payloads on the data shards and atomically publishes references in coordinator etcd. Native coordinator MVCC supplies global revisions, CAS, historical pagination, Watch and lease expiry. Existing backend and legacy modes remain available.

The implementation rejects cross-shard mutations, nested transactions and VALUE comparisons/sorting. It is an explicitly bounded PoC: metadata remains centralized, historical/orphan blobs are not collected, and coordinated backup/restore is not implemented. README, architecture notes and an example explain these limits and the path toward the original sharding goal.

Validation

• Full race-enabled unit and real-etcd integration tests; go vet and workflow lint. Coordinated core statement coverage: 97.1% (234/241).
• Real kube-apiserver 1.35 with mTLS: coordinated 13/13 checks and backend regression 12/12.
• Lost successful data-write/commit replies, concurrent proxies, unavailable shards, historical snapshots, TTL, compaction and restart recovery.
• Workflow runs storage checks on Kubernetes 1.35/1.37 and full kind operations in both modes. Coordinated kind explicitly verifies same-namespace Pods on both shards and LIST-to-WATCH replay through one endpoint with no resource overrides.
• Independent cross-review of implementation, wiring, harness and documentation found no blocking issues. All 7 GitHub checks passed on head 378c5c55d31f531029f0be19e44b41a526f4effc: integration matrix and unit checks.