fix(benchmarks): fix three parsing bugs in parse_results.go and bench_test.go

- parse_results.go: fix metric extraction order — Go outputs custom metrics
  (p50_µs, p95_µs, p99_µs, req/s) BEFORE B/op and allocs/op on the benchmark
  line. The old positional regex had B/op first, so p50/p95/p99 were always
  empty in latency_report.csv. Replaced with separate regexps for each field
  so order no longer matters.

- parse_results.go: remove p95_latency_ms column from throughput_report.csv —
  parallel sweep files only emit ns/op and req/s, never p95 data. The column
  was structurally always empty.

- bench_test.go: remove fmt.Printf from BenchmarkBAPCaller_RPS — the debug
  print raced with Go's own benchmark output line, garbling the result to
  'BenchmarkRPS-N  RPS: N over Ns' which the framework could not parse,
  causing req/s to never appear in the structured output. b.ReportMetric
  alone is sufficient.

benchmarks/e2e/bench_test.go

@@ -1,7 +1,6 @@
 package e2e_bench_test
 
 import (
-	"fmt"
 	"net/http"
 	"sort"
 	"testing"
@@ -162,9 +161,7 @@ func BenchmarkBAPCaller_RPS(b *testing.B) {
 
 	elapsed := time.Since(start).Seconds()
 	if elapsed > 0 {
-		rps := float64(count) / elapsed
-		b.ReportMetric(rps, "req/s")
-		fmt.Printf("  RPS: %.0f over %.1fs\n", rps, elapsed)
+		b.ReportMetric(float64(count)/elapsed, "req/s")
 	}
 }