Use RunParallel for benchmarks

[emcfarlane]

Simplifies the benchmarking to 4 cases, baseline and with interceptors for both unary and streaming cases. Switched to RunParallel to measure overhead per RPC request rather than grouped by the concurrency setting. Reduced streaming to a single round trip to better compare against unary flows:

pkg: connectrpc.com/otelconnect
BenchmarkStreamingBase-8                   19152             58334 ns/op          534279 B/op        204 allocs/op
BenchmarkStreamingWithInterceptor-8        19467             62326 ns/op          546378 B/op        302 allocs/op
BenchmarkUnaryBase-8                       41461             27704 ns/op           14514 B/op        156 allocs/op
BenchmarkUnaryWithInterceptor-8            37642             30383 ns/op           25723 B/op        241 allocs/op
PASS
ok      connectrpc.com/otelconnect      8.435s

Previously:

goos: darwin
goarch: arm64
pkg: connectrpc.com/otelconnect
BenchmarkStreamingServerNoOptions-8                  693           1491135 ns/op         2763098 B/op       1779 allocs/op
BenchmarkStreamingServerClientOption-8               639           1732927 ns/op         3163066 B/op       4751 allocs/op
BenchmarkStreamingServerOption-8                     646           1579154 ns/op         2981825 B/op       3525 allocs/op
BenchmarkStreamingClientOption-8                     688           1515430 ns/op         2918297 B/op       3002 allocs/op
BenchmarkUnaryOtel-8                                5924            198928 ns/op          129339 B/op       1216 allocs/op
2023/10/26 13:37:10 http: TLS handshake error from 127.0.0.1:60015: read tcp 127.0.0.1:60013->127.0.0.1:60015: use of closed network connection
BenchmarkUnary-8                                    6601            183334 ns/op           73278 B/op        791 allocs/op
PASS
ok      connectrpc.com/otelconnect      10.268s

[emcfarlane]

Send triggers 2 response on the sever:

otelconnect-go/pingserver_test.go

Line 56 in a00063c

for i := 0; i < messagesPerRequest; i++ {

so actually need to read twice to drain the service.

Will update the protos to be a pure ping service in a followup PR to make it simpler for testing.