Per-shard atomic disk pressure flags.

When a Bitcask flush fails (e.g., ENOSPC), the shard sets its disk pressure flag. The async write path in Router checks this flag before accepting writes, returning an error instead of silently queuing data that can't be persisted.

The flag is cleared when a flush succeeds, allowing writes to resume automatically once disk space is available (e.g., after compaction reclaims space or an operator adds storage).

Uses :atomics for lock-free ~5ns reads from any process.