prometheus: use a sync.Pool instead of allocating metricdata.ResourceMetrics in Collect (#6472)

Fix #3047 

Use a `sync.Pool` instead of allocating `metricdata.ResourceMetrics` in
`go.opentelemetry.io/otel/exporters/prometheus`

```shell
goos: darwin
goarch: arm64
pkg: go.opentelemetry.io/otel/exporters/prometheus
cpu: Apple M3
               │ before.txt  │             after.txt             │
               │   sec/op    │   sec/op     vs base              │
Collect1-8       8.273µ ± 0%   8.013µ ± 5%       ~ (p=0.065 n=6)
Collect10-8      23.84µ ± 0%   22.40µ ± 3%  -6.02% (p=0.002 n=6)
Collect100-8     148.0µ ± 0%   139.4µ ± 0%  -5.84% (p=0.002 n=6)
Collect1000-8    1.326m ± 0%   1.244m ± 1%  -6.14% (p=0.002 n=6)
Collect10000-8   15.90m ± 1%   14.51m ± 1%  -8.77% (p=0.002 n=6)
geomean          227.9µ        214.3µ       -6.00%

               │  before.txt  │              after.txt              │
               │     B/op     │     B/op      vs base               │
Collect1-8       35.65Ki ± 0%   34.49Ki ± 0%   -3.24% (p=0.002 n=6)
Collect10-8      55.78Ki ± 0%   45.57Ki ± 0%  -18.31% (p=0.002 n=6)
Collect100-8     260.8Ki ± 0%   174.1Ki ± 0%  -33.23% (p=0.002 n=6)
Collect1000-8    2.307Mi ± 0%   1.537Mi ± 1%  -33.38% (p=0.002 n=6)
Collect10000-8   22.47Mi ± 0%   14.75Mi ± 4%  -34.36% (p=0.002 n=6)
geomean          489.8Ki        365.3Ki       -25.42%

               │ before.txt  │             after.txt             │
               │  allocs/op  │  allocs/op   vs base              │
Collect1-8        72.00 ± 0%    66.00 ± 0%  -8.33% (p=0.002 n=6)
Collect10-8       398.0 ± 0%    375.0 ± 0%  -5.78% (p=0.002 n=6)
Collect100-8     3.662k ± 0%   3.491k ± 0%  -4.67% (p=0.002 n=6)
Collect1000-8    36.15k ± 0%   34.62k ± 0%  -4.24% (p=0.002 n=6)
Collect10000-8   361.0k ± 0%   345.7k ± 0%  -4.25% (p=0.002 n=6)
geomean          4.239k        4.008k       -5.47%
```

CHANGELOG.md

@@ -24,6 +24,7 @@ This project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
 
 - Stop percent encoding header environment variables in `go.opentelemetry.io/otel/exporters/otlp/otlplog/otlploggrpc` and `go.opentelemetry.io/otel/exporters/otlp/otlplog/otlploghttp`. (#6392)
 - Ensure the `noopSpan.tracerProvider` method is not inlined in `go.opentelemetry.io/otel/trace` so the `go.opentelemetry.io/auto` instrumentation can instrument non-recording spans. (#6456)
+- Use a `sync.Pool` instead of allocating `metricdata.ResourceMetrics` in `go.opentelemetry.io/otel/exporters/prometheus`. (#6472)
 
 <!-- Released section -->
 <!-- Don't change this section unless doing release -->

exporters/prometheus/exporter.go

@@ -40,7 +40,15 @@ const (
 	spanIDExemplarKey  = "span_id"
 )
 
-var errScopeInvalid = errors.New("invalid scope")
+var (
+	errScopeInvalid = errors.New("invalid scope")
+
+	metricsPool = sync.Pool{
+		New: func() interface{} {
+			return &metricdata.ResourceMetrics{}
+		},
+	}
+)
 
 // Exporter is a Prometheus Exporter that embeds the OTel metric.Reader
 // interface for easy instantiation with a MeterProvider.
@@ -144,9 +152,9 @@ func (c *collector) Describe(ch chan<- *prometheus.Desc) {
 //
 // This method is safe to call concurrently.
 func (c *collector) Collect(ch chan<- prometheus.Metric) {
-	// TODO (#3047): Use a sync.Pool instead of allocating metrics every Collect.
-	metrics := metricdata.ResourceMetrics{}
-	err := c.reader.Collect(context.TODO(), &metrics)
+	metrics := metricsPool.Get().(*metricdata.ResourceMetrics)
+	defer metricsPool.Put(metrics)
+	err := c.reader.Collect(context.TODO(), metrics)
 	if err != nil {
 		if errors.Is(err, metric.ErrReaderShutdown) {
 			return