feat(bench): add Criterion benchmarks and CSIC 2010 dataset converter

8 scanner benchmarks covering allowlist fast path (7.6ns), model path
(172-445ns), and feature extraction (248ns). Python converter script
transforms CSIC 2010 raw HTTP dataset into Sunbeam audit-log JSONL
with realistic scanner feature adaptation.

Signed-off-by: Sienna Meridian Satterwhite <sienna@sunbeam.pt>

.gitignore

@@ -4,3 +4,8 @@ certs/
 *.pem
 *.key
 *.crt
+
+# Training data and model binaries
+*.bin
+*.jsonl
+heuristics.toml