feat: adaptive trace sampling based on latency and error status

[mayankpande88]

Problem

The trace sampling rate is a static float (default 1.0). On a busy node, a service doing 10k req/s generates 10k spans/s regardless of whether they're interesting.

Proposed Behavior

• Always sample: Error responses (5xx, connection failures), slow requests (> P95)
• Normal sampling: Success responses at configured rate
• Reduced sampling: Health checks, readiness probes (if not already filtered)

This ensures operators always see the interesting traces while reducing volume for the common case.

Implementation

Add a SmartSampler that wraps the existing shouldSample():

func shouldSample(status int, duration time.Duration) bool {
    if status >= 500 || duration > slowThreshold {
        return true // always sample errors and slow requests
    }
    return rand.Float64() < samplingRate
}

The slow threshold could be dynamically computed from a rolling P95 or configured via flag.

Impact

Reduces OTel collector load while preserving signal quality.