fix: [listener] make observedConnection.Close idempotent to prevent active connection metric negative drift

[thinker0]

Description

This pull request solves the negative drift bug on the trickster_proxy_active_connections Prometheus gauge metric.

Root cause:
In pkg/proxy/listener/listener.go, the Close() method on observedConnection was not idempotent. Go's net/http server can invoke Close() multiple times in different execution paths (e.g., Serve goroutine defer, timeout handlers, Hijack). Each invocation decremented the active connections gauge (metrics.ProxyActiveConnections.Dec()) and incremented the closed total, leading to monotonic negative drift over time under normal load.

Solution:
Applied sync.Once inside observedConnection to guarantee that connection close metrics are updated exactly once per connection lifecycle, ensuring gauge consistency.

Also added high-stress unit testing (TestObservedConnectionIdempotentClose in listener_test.go) under go test -race to verify idempotence and prevent regressions.

Type of Change

• Bug fix
• New feature
• Optimization
• Test coverage
• Documentation
• Infrastructure

AI Disclosure

• This contribution DOES include AI-generated changes, and I have reviewed the relevant contributing guidelines.

[crandles]

a second close should return an err

Suggested change

	o.once.Do(func() {
	metrics.ProxyActiveConnections.Dec()
	metrics.ProxyConnectionClosed.Inc()
	})
	if err == nil {
	metrics.ProxyActiveConnections.Dec()
	metrics.ProxyConnectionClosed.Inc()
	}

I think it would be simpler to skip sync.Once here if we add an err check

[crandles]

you're not checking for success here, unlike what the comment says