feat(hotblocks): expose RocksDB stall metrics

[mo4islona]

Summary

• expose bounded RocksDB write-stall, compaction, flush, memtable, snapshot, and background-error metrics
• add low-volume startup phase logs around DB open, dataset controller initialization, and disk recovery
• verify selected properties remain available with RocksDB statistics disabled
• include an ignored reproducible benchmark for scrape overhead on reads and writes

Why

When all datasets stop advancing together, these signals distinguish a RocksDB hard or soft write stall from blocking-pool or upstream pipeline starvation. Per-column-family backlog metrics make the responsible LSM area visible without enabling expensive RocksDB statistics.

Performance

The collector reads fixed-cardinality in-memory integer properties only during Prometheus scrapes. Expensive properties that iterate SST files or versions are excluded.

Release benchmark under continuous scraping, much harsher than a production scrape interval:

• full scrape p50: 11.166 us
• write p50: 9.125 us baseline, 8.667 us with scraping
• read p50: 1.166 us baseline, 1.250 us with scraping

Verification

• cargo test -p sqd-hotblocks
• cargo test -p sqd-storage
• cargo +nightly fmt --all -- --check
• cargo clippy --all-targets --all-features -- -D clippy::correctness
• git diff --check