Add integrated, scaling, and per-record benchmarks; correct README perf claims

[eitamring]

Benchmark and docs

• New benchmarks: integrated text pipeline (source -> rules -> JSONL sink), structured pipeline, sink to io.Discard vs file, per-record engine allocations across worker counts, ruleset scaling at 20/100/1000 rules, middle-hit, and the unprefilterable-regex miss.
• Rules benchmarks now check errors.Is instead of comparing error strings; ruleset construction moved out of the timed region.
• README: the worst-case row is relabeled as the benchmark ruleset's 20-rule regex-heavy miss, with a note that patterns without an extractable literal pay full regex cost (~101 us at 20 rules, reproducible via the named benchmark); added the integrated-pipeline row so the discard-sink number is not mistaken for what a CLI run pays.

make bench covers all new benchmarks and completes in ~75s.