feat: Early load ustr (#792)

## Problem
The first time `Ustr::from()` is called, it takes 10+ ms to init. Right
now this is on the critical path and adds to cold start time.

## This PR
Start it early in a separate thread, so when it is needed later, it will
be already initialized.

## Result
Using `extensiononly` stack on Node 22:

### Before
- `Datadog Next-Gen Extension ready in`: **57.2** ±3.1 ms (95%
confidence interval, 17 data points)
- `Metrics aggregator created in`: **10+ ms** (10,000+ us)

### After
- `Datadog Next-Gen Extension ready in`: **49.7**±1.9 ms (95% confidence
level, 10 data points)
- `Metrics aggregator created in`: **277±186 us** (95% confidence level,
10 data points)

Thanks @duncanista for identifying this performance bottleneck!
Jira: https://datadoghq.atlassian.net/browse/SVLS-7456

Author: lym953

bottlecap/Cargo.lock

@@ -42,6 +42,7 @@ opentelemetry-proto = { version = "0.29", features = ["trace", "with-serde", "ge
 opentelemetry-semantic-conventions = { version = "0.29", features = ["semconv_experimental"] }
 rustls-native-certs = { version = "0.8.1", optional = true }
 axum = { version = "0.8.4", default-features = false, features = ["default"] }
+ustr = { version = "1.0.0", default-features = false }
 # If you are adding or updating a datadog-owned code dependency, please ensure
 # that it has a clippy.toml rule for disallowing the reqwest::Client::builder
 # method in favor of our

bottlecap/Cargo.toml

@@ -90,6 +90,7 @@ use tokio::{sync::Mutex as TokioMutex, sync::RwLock, sync::mpsc::Sender, task::J
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, error};
 use tracing_subscriber::EnvFilter;
+use ustr::Ustr;
 
 #[allow(clippy::struct_field_names)]
 struct PendingFlushHandles {
@@ -331,6 +332,7 @@ async fn register(client: &Client) -> Result<RegisterResponse> {
 #[tokio::main]
 async fn main() -> Result<()> {
     let start_time = Instant::now();
+    init_ustr();
     let (aws_config, aws_credentials, config) = load_configs(start_time);
 
     enable_logging_subsystem(&config);
@@ -382,6 +384,14 @@ async fn main() -> Result<()> {
     }
 }
 
+// Ustr initialization can take 10+ ms.
+// Start it early in a separate thread so it won't become a bottleneck later when SortedTags::parse() is called.
+fn init_ustr() {
+    tokio::spawn(async {
+        Ustr::from("");
+    });
+}
+
 fn load_configs(start_time: Instant) -> (AwsConfig, AwsCredentials, Arc<Config>) {
     // First load the AWS configuration
     let aws_config = AwsConfig::from_env(start_time);
@@ -479,13 +489,21 @@ async fn extension_loop_active(
         event_bus.get_sender_copy(),
     );
 
+    let metrics_aggr_init_start_time = Instant::now();
     let metrics_aggr = Arc::new(Mutex::new(
         MetricsAggregator::new(
             SortedTags::parse(&tags_provider.get_tags_string()).unwrap_or(EMPTY_TAGS),
             CONTEXTS,
         )
         .expect("failed to create aggregator"),
     ));
+    debug!(
+        "Metrics aggregator created in {:} microseconds",
+        metrics_aggr_init_start_time
+            .elapsed()
+            .as_micros()
+            .to_string()
+    );
 
     let metrics_flushers = Arc::new(TokioMutex::new(start_metrics_flushers(
         Arc::clone(&api_key_factory),