fullhistory: add bench-ingest subcommand for ingestion benchmarks (#771)

[marwen-abid]

What

Adds stellar-rpc bench-ingest cold|hot: benchmarks for the production
ingestion paths (ingest.WriteColdChunk, ingest.HotService). A CSV
MetricSink aggregates the signals the engine already emits into
per-stage percentile reports (one CSV per cold data type, hot.csv,
driver.csv) — no re-instrumentation, the engine under measurement is
the one that ships.

• Sources: frozen local packs (--source=pack) or the backfill BSB
  backend (--source=bsb) against GCS or S3; cold runs N chunks with
  --chunk-workers fan-out, hot caps runs with --num-ledgers.
• Guards: takes the daemon's exclusive flocks on every write root,
  refuses in-place repacks and pre-existing hot DBs, bounds --chunk
  at the last valid chunk ID. Failed/interrupted runs persist PARTIAL
  CSVs for completed work.
• Supporting changes: ingest exports DataType*/Stage* label
  constants, fhtest.EventLCMBytes is the canonical event-bearing LCM
  fixture, geometry.LedgerPackPath owns the pack-path formula.

Example run (real pubnet data from the public S3 lake):

    stellar-rpc bench-ingest cold --source=bsb --datastore-type=S3 \
        --region=us-east-2 \
        --bucket-path=aws-public-blockchain/v1.1/stellar/ledgers/pubnet \
        --chunk=6100 --num-chunks=1 --types=events \
        --cold-out-dir=/tmp/cold --out=/tmp/bench-out

Smoke-tested on pubnet chunk 6100: cold events = 6.15M events, 14.2s of
ingest inside 46.8s wall (rest is S3 download); hot commit fsync
dominates at p50 7ms/ledger, as designed.

Note for CSV consumers: read_blocked's n is ledgers − 1 — the first
pull (source setup) is excluded so max/total reflect steady-state waits.

Why

We need baseline ingestion numbers for the full-history storage engine —
per-stage cold throughput and hot per-ledger latency — to tune it
(#836), to run the apply-load measurement campaign, and to seed the
bench-in-CI work (#775). Benchmarking the shipped entry points from a
subcommand keeps the numbers honest: what's measured is exactly what
runs in production.

Query benchmarks are split out to #856 (closes #771).