perf(mempool): Replace sync.Mutex with sync.Once (#13293)

`sync.Once` uses an atomic integer as gate to check if the slab is already initialized and therefore yields a  ~9% performance improvement for `get` operations.

```
goos: linux
goarch: amd64
pkg: github.com/grafana/loki/v3/pkg/util/mempool
cpu: 11th Gen Intel(R) Core(TM) i7-1185G7 @ 3.00GHz
             │  bench.old  │             bench.new              │
             │   sec/op    │   sec/op     vs base               │
Slab/1KB-8     138.5n ± 2%   125.1n ± 2%  -9.71% (p=0.000 n=10)
Slab/1MB-8     140.5n ± 3%   128.1n ± 2%  -8.83% (p=0.000 n=10)
Slab/128MB-8   143.0n ± 3%   130.4n ± 3%  -8.81% (p=0.000 n=10)
geomean        140.7n        127.8n       -9.12%
```

Signed-off-by: Christian Haudum <christian.haudum@gmail.com>

pkg/util/mempool/pool.go

@@ -20,7 +20,7 @@ var (
 type slab struct {
 	buffer      chan unsafe.Pointer
 	size, count int
-	mtx         sync.Mutex
+	once        sync.Once
 	metrics     *metrics
 	name        string
 }
@@ -49,11 +49,7 @@ func (s *slab) init() {
 
 func (s *slab) get(size int) ([]byte, error) {
 	s.metrics.accesses.WithLabelValues(s.name, opTypeGet).Inc()
-	s.mtx.Lock()
-	if s.buffer == nil {
-		s.init()
-	}
-	defer s.mtx.Unlock()
+	s.once.Do(s.init)
 
 	// wait for available buffer on channel
 	var buf []byte

pkg/util/mempool/pool_test.go

@@ -124,7 +124,6 @@ func BenchmarkSlab(b *testing.B) {
 	} {
 		b.Run(flagext.ByteSize(uint64(sz)).String(), func(b *testing.B) {
 			slab := newSlab(sz, 1, newMetrics(nil, "test"))
-			slab.init()
 			b.ResetTimer()
 
 			for i := 0; i < b.N; i++ {