stream: move serverStream.mu to cache line 2 to eliminate false sharing

serverHeaderBinlogged (written unsynchronised in SendHeader and Send) and
mu shared cache line 3 (192–255) after the initial tail-grouping commit.
Concurrent writes to serverHeaderBinlogged invalidate the cache line before
every mu.Lock(), adding ~5–21× overhead under concurrent load.

Moving mu to immediately before trInfo places it at offset 184 (cache line 2:
128–191).  The tail bools remain at offset 240+ (cache line 3), so stressor
writes no longer affect mu.Lock() latency.

Benchmark before this commit (serverHeaderBinlogged stressor):
  stressors=0   3.9 ns
  stressors=1  20.1 ns  (5.2×)
  stressors=4  79.9 ns (20.4×)

Benchmark after (different cache lines):
  stressors=0  3.7 ns
  stressors=1  3.8 ns  (~1×)
  stressors=4  3.7 ns  (~1×)

No size change: serverStream remains 248 B.

Fixes #9349

Author: gidotencate

stream.go

@@ -1728,20 +1728,20 @@ type serverStream struct {
 
 	maxReceiveMessageSize int
 	maxSendMessageSize    int
-	trInfo                *traceInfo
+
+	// mu guards trInfo.tr after the service handler runs.
+	mu     sync.Mutex
+	trInfo *traceInfo
 
 	statsHandler stats.Handler
 
 	binlogs []binarylog.MethodLogger
 
-	mu sync.Mutex // protects trInfo.tr after the service handler runs.
-
 	// Bool fields are grouped at the tail to eliminate the alignment padding
 	// that would otherwise follow each bool when the next field is pointer- or
 	// int-sized. See https://github.com/grpc/grpc-go/issues/9349 for benchmarks.
 	// Add new bool fields here, not inline above.
 
-	// Not guarded by mu.
 	recvFirstMsg bool // set after the first message is received
 	// serverHeaderBinlogged indicates whether server header has been logged. It
 	// will happen when one of the following two happens: stream.SendHeader(),