Add serving endpoint metrics export with Prometheus parser

[jralfonsog]

Summary

• Metrics export: export_serving_endpoint_metrics() wraps the SDK's export_metrics API, returning both parsed structured dicts and raw Prometheus text
• Prometheus parser: _parse_prometheus_metrics() handles the full OpenMetrics exposition format — HELP/TYPE metadata, labels, histogram buckets, and optional timestamps
• Structured output: Each metric becomes {"name", "labels", "value", "help", "type"} so LLMs can reason about endpoint health without parsing raw text

Available metrics

Metric	Type	Description
cpu_usage_percentage	gauge	CPU utilization across replicas
mem_usage_percentage	gauge	Memory utilization across replicas
request_count_total	gauge	Requests processed per minute
request_4xx_count_total	gauge	Client error rate
request_5xx_count_total	gauge	Server error rate
request_latency_ms	histogram	Round-trip latency (P50/P99 via buckets)
gpu_usage_percentage	gauge	GPU utilization (GPU workloads only)
gpu_memory_usage_percentage	gauge	GPU memory (GPU workloads only)

Changes

Layer	File	What changed
Core	serving/endpoints.py	+export_serving_endpoint_metrics(), +_parse_prometheus_metrics(), added re import and ResourceDoesNotExist/NotFound for typed error handling
Core	serving/__init__.py	Export new function
MCP	tools/serving.py	+1 @mcp.tool(timeout=30) wrapper
Tests	tests/unit/test_serving_metrics.py	9 unit tests: parser (gauge, counter, histogram, empty, no-labels, timestamps) + export (success, not-found, empty)

Test plan

• 9 unit tests pass (parser + export function)
• Ruff lint + format pass (line-length=120, py311)
• Integration tested against aws-fe workspace — returned 3 metrics for a scaled-to-zero sklearn endpoint
• Prometheus timestamp handling verified (real API returns timestamps, parser strips them correctly)
• Not-found returns clean error via typed SDK exceptions

This pull request was AI-assisted by Isaac