async sub-block backfill for forked writes

[jaredLunde]

• When a forked VM writes a sub-block (e.g., 4KB) to a 128KB block that only exists in S3, the write now lands on local SSD immediately instead of synchronously fetching the full block from S3
• Tracks unfilled sub-regions via a 32-bit bitmap in partial_blocks: DashMap<usize, AtomicU32>, with background tasks backfilling from S3
• Extends WAL with 24-byte partial entries (high-bit flag + bitmap) so partial state survives crashes
• Flush skips partial blocks (no garbage upload); on-demand reads merge SSD + S3 data
• Post-recovery backfill spawning ensures partial blocks don't stay dirty indefinitely
• Cap of 1024 simultaneous partial blocks; beyond that, falls back to synchronous fetch
• Adds a ton more data integrity tests