[Resume tracker] Add resume-path memory telemetry and regression soak test

[aivong-openhands]

Parent tracker

• [Tracker] Make sandbox resume idempotent and eliminate repeated allocation churn #244

Goal

Make resume-path allocation and retention observable, and prevent regressions with a repeated-request soak test.

Scope

• Add low-cardinality counters for resume outcomes: resumed, already_active, conflict, not_found, rate_limited, and error.
• Export process RSS and Python heap metrics separately.
• Export GC generation/collection counts and active asyncio task count.
• Add HTTP client connection/pool and trace-writer queue/drop telemetry where available.
• Replace request-frequency full tracebacks for expected conflicts with sampled warnings and counters.
• Build a soak test that repeatedly exercises no-op resume requests and records RSS/heap growth.
• Define an acceptable memory-growth threshold for CI or a dedicated performance environment.

Acceptance criteria

• Operators can distinguish Python heap growth from allocator/RSS high-water behavior.
• Expected resume conflicts are measurable without high-volume traceback logging.
• The soak test fails when memory grows beyond the documented threshold.
• Metrics do not use user, sandbox, runtime, or client identifiers as tags.

This issue was created by an AI agent (OpenHands) on behalf of the user.