tests: refactor benchmarks to remove test server overhead

[costela]

The httptest Server introduces unrelated overhead to the benchmarks.
Ideally we'd only measure the impact on the wrapped handler.

Related to #19

[felixge]

🙏 Thanks. I think I originally put the test server into the benchmark so that the overhead created by httpsnoop is "put into perspective". After applying your patch people might think that httpsnoop is causing 240x overhead 🙈. Anyway, I agree that it's more useful to exclude the test server overhead while trying to optimize things. So I'm okay with this change.

$ go test -bench .
goos: darwin
goarch: amd64
pkg: github.com/felixge/httpsnoop
cpu: Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz
BenchmarkBaseline-12               	809914038	        1.479 ns/op	      0 B/op	      0 allocs/op
BenchmarkCaptureMetrics-12         	3394172	      353.9 ns/op	    225 B/op	      7 allocs/op
BenchmarkCaptureMetricsTwice-12    	1715709	      698.2 ns/op	    450 B/op	     14 allocs/op
BenchmarkWrap-12                   	8871846	      133.1 ns/op
PASS
ok  	github.com/felixge/httpsnoop	6.244s

[costela]

Yes, that's a very good point. But I'd argue that this perspective is best given in a README, if these results are mentioned somewhere. If we're running the benchmarks while working on the code, this added "context" is mostly implicit?