perf(despeckle): skip the metrics-only component-counting passes

[P4suta]

Stacked on #31 (uses its benchCleaner baseline for the A/B).

Why

clean() counted 8-connected components before and after every page — two full connected-component labelings, 13.5% of the page per the committed baseline — yet the counts feed nothing but the HTML report's charts and a summary log line. The pdfbook pipeline never sets a report dir, so its hot path paid for numbers nobody read.

What

• ProcessResult: component counts become OptionalInt; a counted 4-int convenience constructor keeps every existing construction compiling unchanged; componentsRemoved() stays present-or-zero int (documented); new withoutComponentStats(...) factory, hasComponentStats(), blackPixelsRemoved().
• ProcessOptions: 6th knob collectComponentStats (default true via the delegating 5-knob constructor — all existing callers unchanged) + withoutComponentStats() wither carried through withDpi.
• LeptonicaPageCleaner: both counting passes guarded; the black-pixel counts still run unconditionally (they feed the 3% over-removal guard).
• DespeckleService: counting is on exactly when reportDir != null — the report path (HTML charts, totals, the "N component(s) removed" line) is byte-identical. Without a report, the summary logs in always-measured black-pixel terms; PdfBatchService drops the would-always-be-0 component clause without reports.

Measured

metric	before	after
clean() single-threaded	161.8ms	139.5ms (−13.8%) — matches the baseline's countConnComp share
despeckle stage (-j8, 200p)	10.19s	9.57s (−6.1%) — partially absorbed by memory-bandwidth saturation
conv (-j8)	14.23s	13.61s (−4.4%)

The −4.4% conv alone sits just under the 5% gate; it's a true work reduction (not a tradeoff) and compounds with the DWA morphology PR that follows.

Verification

• Full ./gradlew check green; zero edits to existing test files' assertions (the compat constructors carry them).
• New tests: domain (factory/wither/flag carry), infrastructure (counting-off clean still cleans pixel-identically and the over-removal guard works), application (capturing fake pins reportDir↔flag in both directions).

🤖 Generated with Claude Code

Re-file of #32 (same commit, cherry-picked onto main).