Add TypedCache[T] for pre-deserialized tree cache

[epk]

Summary

Adds TypedCache[T], a generic wrapper around TreeCache that deserializes node data once on write (ZK event) and serves concrete typed objects on read. This eliminates per-access Unmarshal calls.

API

// Create a typed cache wrapping an existing TreeCache.
cache := zk.NewTreeCache(conn, "/mcs.v1", zk.WithTreeCacheIncludeData(true))
typed := zk.NewTypedCache[mcsv1.Metadata](cache, json.Unmarshal,
    zk.WithTypedCacheListener[mcsv1.Metadata](metricsListener))

// Read pre-deserialized objects (lock-free).
val, stat, err := typed.GetTyped("/services/foo")

// Iterate all typed entries.
for path, val, stat := range typed.AllTyped() { ... }

// Count entries.
n := typed.Len()

Design

• Thread safety: atomic.Pointer[sync.Map] gives lock-free concurrent reads. On re-sync after reconnection, a fresh map is built by walking the TreeCache and swapped atomically — no reader ever sees partial state.
• Listener chaining: TypedCache implements TreeCacheListener and installs itself on the TreeCache. WithTypedCacheListener chains an inner listener that receives all events after TypedCache processes them.
• Non-deserializable nodes: Silently skipped (logged at debug level). GetTyped returns ErrNoNode for them.
• Path handling: Works with both relative and absolute path modes (WithTreeCacheAbsolutePaths).

Tests

• Initial sync populates typed entries
• Node create/update/delete reflected in GetTyped
• Re-sync after disconnect clears stale entries
• Inner listener chaining works
• Absolute paths mode works
• Non-deserializable nodes silently skipped
• Concurrent reads during sync (safe with -race)

[jerr]

I'm not a zk expert, so I might have missed something, but this PR makes sense and should be a good improvement. 👍