ethdb/leveldb: fix metrics double-differencing; use per-interval DBStats directly

[MozirDmitriy]

The LevelDB metrics loop subtracted previous samples from DBStats fields that are already per-interval (Stats resets internal buffers on each call). This caused double-differencing and under-reported/zero deltas.

Changes:

• Mark compaction meters using current interval values from LevelDB Stats (no previous-sample subtraction).
• Mark write-delay meters with current WriteDelayCount/Duration directly; adjust stall warning to check zero current-interval deltas while paused.
• Mark disk IO meters with current IORead/IOWrite interval values.
• Keep disk size as an absolute gauge; remove previous-sample buffers and related state.

This aligns LevelDB metrics with its Stats semantics and matches Pebble’s approach (which uses cumulative counters, hence still needs deltas).

[rjl493456442]

I don't think this fix is correct. What we actually care about is the statistical difference
between intervals, which should be uploaded to the meters.

Besides, since we plan to deprecate go-leveldb, it’s not worthwhile to accept changes against it.