feat(bench): add fusio io probes to compaction benchmark

[belveryin]

Closes #586

Summary

Add benchmark-only Fusio IO probes for compaction/read amplification and emit the counters in the compaction benchmark JSON artifact.

What changed

• Added benchmark-only probed FS wrappers (ProbedLocalFs / ProbedLocalFile) to count:
  • read_ops
  • write_ops
  • bytes_read
  • bytes_written
  • ssts_touched (best effort; unique .parquet paths under /sst/)
• Integrated counters into artifact schema:
  • scenarios[].setup.io
  • scenarios[].summary.io
• Reset counters per scenario measurement to avoid cross-scenario leakage.
• Added DbBuilder::on_durable_fs(...) so the benchmark can inject a wrapped durable FS without changing normal production behavior.
• Bumped benchmark artifact schema version to 2.

Validation

• cargo +nightly fmt --all
• cargo clippy -- -D warnings
• cargo test
• cargo bench --bench compaction_local

[ethe]

Is read_post_compaction intended to mean "first compaction observed" or "compaction quiesced"?

[ethe]

read_post_compaction can start measuring before compaction fully settles, so results may be noisy/inconsistent.

[ethe]

read_baseline/read_post_compaction can panic on valid low-ingest env configs (list_versions returned no versions), which breaks the "env-only tunable benchmark" expectation.

[ethe]

read_post_compaction can start measuring before compaction fully settles, so results may be noisy/inconsistent.

[belveryin]

• Split the old read_post_compaction into two explicit scenarios: read_after_first_compaction_observed and read_compaction_quiesced, with separate wait logic for each.
• Made low-ingest / no-SST setup non-fatal by skipping inapplicable scenarios with a clear log message instead of panicking.
• Added strict config validation so TONBO_COMPACTION_BENCH_INGEST_BATCHES must be > 0 and fails fast with an explicit error.