Add cheap load tests for concurrent conversations

Problem

A regression recently prevented LLM calls from running in parallel, effectively making the agent server single-threaded whenever LLM calls were moderately slow. We do not currently have a cheap load test that would catch this behavior across concurrent conversations.

Related context:

• OpenHands/OpenHands#16459
• OpenHands/software-agent-sdk#4473

Proposed test

Add a deterministic concurrent-conversation load test for the agent server/SDK that:

• Starts multiple conversations at the same time.
• Uses a cheap local/fake LLM implementation with a controlled delay, avoiding paid external LLM calls.
• Verifies LLM requests overlap rather than execute serially.
• Records or asserts aggregate wall-clock time/throughput with enough tolerance to avoid CI flakiness.
• Exercises the real conversation and LLM request path involved in production, rather than testing only an isolated helper.
• Produces useful diagnostics when concurrency regresses.

Acceptance criteria

• A repeatable test runs several conversations concurrently through the real agent-server/SDK path.
• The test fails when LLM calls are unintentionally serialized.
• The test requires no paid model credentials and is inexpensive enough for regular CI execution.
• Timing thresholds are documented and tolerant of normal CI variance.
• The test is included in an appropriate CI workflow.

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

Start by tracing the real agent-server/SDK conversation and LLM request path described in the issue, then identify the appropriate CI workflow. Add a deterministic test using a cheap local or fake LLM with controlled delay, and verify it detects serialized calls while remaining tolerant of normal CI variance and producing useful diagnostics.