feat(fusio-manifest): add run-v2 sparse index and block cache

[ethe]

Why / Intent

fusio-manifest point-get on merge-tree runs could still degrade into reading large run payloads and scanning linearly.
This is especially costly for object storage (e.g. S3), where unnecessary range reads amplify latency and I/O.
This PR changes the read path to index-first + block-level reads:

• filter and locate with run index metadata
• read only the minimal candidate block
• cache hot blocks to avoid repeated remote reads
  Backward compatibility is intentionally not preserved in this change.

Benefits

Measured on local-disk benchmark with cache disabled, bloom+sparse/block path significantly reduces wasted I/O in in-range misses:

• L1-InRangeMiss
  • ckpt_payload_range: 980 -> 9
  • requested_bytes: 9,416,820 -> 86,481 (~-99.1%)
  • ops/s: 3204.47 -> 3702.02 (+15.5%)
  • p50: 312.17us -> 266.58us (-14.6%)
• MultiLevel-InRangeMiss
  • ckpt_payload_range: 988 -> 6
  • requested_bytes: 9,746,620 -> 59,190 (~-99.4%)
  • ops/s: 2535.56 -> 2668.60 (+5.2%)
  • p50: 394.29us -> 373.12us (-5.4%)
    Net effect: much lower worst-path point-get cost through aggressive run skipping and minimal block reads.

Compatibility

This PR intentionally introduces a format break:

• no backward compatibility layer for old run/checkpoint payloads
• migration requires regenerating runs (re-compaction / rebuild)

Validation

• cargo +nightly fmt --all -- --check
• cargo clippy -p fusio-manifest -- -D warnings
• cargo test -p fusio-manifest
• cargo check -p fusio-manifest --benches --features cache-moka