TUN-3208: Reduce copies and allocations on h2mux write path. Pre-allocate 16KB write buffer on the first write if possible. Use explicit byte array for chunks on write thread to avoid copying through intermediate buffer due to io.CopyN.

benchmark                                    old ns/op     new ns/op     delta
BenchmarkSingleStreamLargeResponseBody-8     17786594      12163494      -31.61%

benchmark                                    old allocs     new allocs     delta
BenchmarkSingleStreamLargeResponseBody-8     17086          15869          -7.12%

benchmark                                    old bytes     new bytes     delta
BenchmarkSingleStreamLargeResponseBody-8     58215169      21604391      -62.89%

h2mux/h2mux.go

@@ -7,11 +7,12 @@ import (
 	"sync"
 	"time"
 
-	"github.com/cloudflare/cloudflared/logger"
 	"github.com/prometheus/client_golang/prometheus"
 	"golang.org/x/net/http2"
 	"golang.org/x/net/http2/hpack"
 	"golang.org/x/sync/errgroup"
+
+	"github.com/cloudflare/cloudflared/logger"
 )
 
 const (
@@ -21,6 +22,7 @@ const (
 	defaultTimeout           time.Duration = 5 * time.Second
 	defaultRetries           uint64        = 5
 	defaultWriteBufferMaxLen int           = 1024 * 1024 // 1mb
+	writeBufferInitialSize   int           = 16 * 1024 // 16KB
 
 	SettingMuxerMagic http2.SettingID = 0x42db
 	MuxerMagicOrigin  uint32          = 0xa2e43c8b
@@ -206,9 +208,9 @@ func Handshake(
 		initialStreamWindow:     m.config.DefaultWindowSize,
 		streamWindowMax:         m.config.MaxWindowSize,
 		streamWriteBufferMaxLen: m.config.StreamWriteBufferMaxLen,
-		r:              m.r,
-		metricsUpdater: m.muxMetricsUpdater,
-		bytesRead:      inBoundCounter,
+		r:                       m.r,
+		metricsUpdater:          m.muxMetricsUpdater,
+		bytesRead:               inBoundCounter,
 	}
 	m.muxWriter = &MuxWriter{
 		f:               m.f,