Reduce memory consumption during proxying

[simonmittag]

Why

During this run, we consumed 8GB for 64k connections on HTTP1.1: https://github.com/simonmittag/j8a-performance/blob/master/results/7/pt-j8a1/output.json

{"level":"info","srvId":"e21fbe1a","pid":10468,"pidCpuCorePct":"788.39","pidMemPct":"9.28","pidDwnOpenTcpConns":65536,"pidDwnMaxOpenTcpConns":65536,"pidUpOpenTcpConns":34188,"pidUpMaxOpenTcpConns":34188,"pidMemRssBytes":1990791168,"pidMemVmsBytes":2714484736,"pidMemSwapBytes":0,"pidOSThreads":36,"pidOSUlimit":1048576,"time":"2022-06-20T00:32:01Z","message":"server performance"}
{"level":"info","srvId":"e21fbe1a","pid":10468,"pidCpuCorePct":"793.44","pidMemPct":"38.14","pidDwnOpenTcpConns":65536,"pidDwnMaxOpenTcpConns":65536,"pidUpOpenTcpConns":31660,"pidUpMaxOpenTcpConns":34188,"pidMemRssBytes":8177192960,"pidMemVmsBytes":8857845760,"pidMemSwapBytes":0,"pidOSThreads":36,"pidOSUlimit":1048576,"time":"2022-06-20T00:33:06Z","message":"server performance"}
{"level":"info","srvId":"e21fbe1a","pid":10468,"pidCpuCorePct":"788.87","pidMemPct":"38.92","pidDwnOpenTcpConns":65536,"pidDwnMaxOpenTcpConns":65536,"pidUpOpenTcpConns":31965,"pidUpMaxOpenTcpConns":34188,"pidMemRssBytes":8345534464,"pidMemVmsBytes":9135472640,"pidMemSwapBytes":0,"pidOSThreads":36,"pidOSUlimit":1048576,"time":"2022-06-20T00:34:11Z","message":"server performance"}
{"level":"info","srvId":"e21fbe1a","pid":10468,"pidCpuCorePct":"786.43","pidMemPct":"37.28","pidDwnOpenTcpConns":46440,"pidDwnMaxOpenTcpConns":65536,"pidUpOpenTcpConns":29012,"pidUpMaxOpenTcpConns":34188,"pidMemRssBytes":7993102336,"pidMemVmsBytes":9275465728,"pidMemSwapBytes":0,"pidOSThreads":36,"pidOSUlimit":1048576,"time":"2022-06-20T00:35:14Z","message":"server performance"}

What

• reduce allocations for proxying, especially body reads.
• can we get away with not copying the body if we want to do retries? would have to see if for retries we can re-read the body
• for response rewriting this should be ok, except we'd have to pipe what is being read through Gzip and Brotli encoders as necessary. if we do that, it'd have to be buffered to minimum blocksize for both algs