feat: support ADK 2.13 info efficiency metrics and bump Python deps - #265

Supersedes #262 and fixes what it broke.

ADK 2.13 adds an INFORMATIONAL eval status and four efficiency metrics: tool_call_count_v1, inference_call_count_v1, token_usage_v1 and invocation_duration_v1. They appeared in the metric catalog but failed at runtime, because agentevals always sent a threshold.

• Adds EVAL_STATUS_INFORMATIONAL to the v1alpha1 EvalStatus enum (additive, wire compatible) and maps it in the gRPC converter, CLI output ([INFO]) and OTel evaluation events (informational label).
• The four metrics take no threshold: a per metric threshold is rejected with a clear error, and run level thresholds (CLI, MCP, UI) skip them.
• Fills the fields these metrics read when invocations are rebuilt from traces: one model call event per LLM call with token usage, plus turn duration split into model, tool and framework time. Model time comes from the inference span itself, so wrapper spans that also cover tool execution (ADK call_llm) don't hide tool time.
• UI shows informational results as neutral values and leaves them out of pass rate, average score and pass/fail badges.

Install changes

The google-adk minimum is now 2.13, and the protobuf upper bound moves from <7 to <8.