test(mcp): add runner-level coverage for upstream tool errors

[OiPunk]

Summary

This PR adds runner-level regression coverage for MCP upstream tool errors (issue #879).

Problem

Issue #879 reports that when an MCP upstream service returns an error (e.g., HTTP 422), the run can abort with an AgentsException instead of returning model-visible tool output.

Current code already implements graceful handling via failure_error_function in MCP tool wrappers, but there was no runner-level test that locks in this behavior end-to-end.

Changes

• Added test_runner_handles_mcp_upstream_errors_as_tool_output:
  • Simulates an MCP server whose call_tool raises an upstream error.
  • Verifies Runner.run and Runner.run_streamed both continue and complete with the next model turn.
• Added test_runner_raises_when_mcp_failure_error_function_disabled:
  • Verifies explicit mcp_config={"failure_error_function": None} preserves raising behavior.

Why this is useful

• Prevents regressions on a real production concern: MCP upstream errors should not unexpectedly terminate runs under default configuration.
• Clarifies behavior boundaries between default error handling and explicit opt-out.

Local validation

• uv run --with pytest pytest -q tests/mcp/test_runner_calls_mcp.py
• uv run --with ruff ruff check tests/mcp/test_runner_calls_mcp.py
• uv run --with pytest pytest -q tests/mcp/test_mcp_util.py -k 'graceful_error_handling or failure_error_function'

Closes #879

[OiPunk]

Follow-up update:

• The initial CI typecheck failure was caused by a mypy inference mismatch in the new test (RunResultStreaming vs RunResult variable reuse across branches).
• I pushed c99c267a to split branch-local variables (streamed_result / non_streamed_result) without changing test behavior.

Revalidated locally:

• uv run --with mypy mypy tests/mcp/test_runner_calls_mcp.py
• uv run --with pytest pytest -q tests/mcp/test_runner_calls_mcp.py

Both pass.