opentelemetry-util-genai: emit the per-operation execute_tool and invoke_agent metrics

What problem do you want to solve?

InvocationMetricsRecorder.record()
sends every invocation type to a single histogram, gen_ai.client.operation.duration.
Only WorkflowInvocation was given its own instrument; tool and agent invocations still
land in the client-operation metric, separated only by gen_ai.operation.name.
Four semconv instruments are missing:

All are Recommended, the highest level semconv defines for signals. Their attribute
sets also differ from the client metric's — gen_ai.execute_tool.duration requires
gen_ai.tool.name and carries no gen_ai.operation.name — so slicing one histogram by
operation name is not equivalent.
Two smaller conformance gaps in the existing workflow metric, if worth folding in:

• it reuses the client bucket advisory (0.01 … 81.92) instead of its own (1 … 7200)
• it sets gen_ai.operation.name, which is not in its attribute table

Describe the solution you'd like

• Add the four instruments with their own bucket advisories.
• Let each invocation declare its duration instrument and build metric attributes from
  that metric's table, instead of adding another record_<type>() per operation.
• Scope inference_calls / tool_calls to the agent that issued them, so sub-agent
  calls are counted once.

Describe alternatives you've considered

No response

Additional Context

No response

Would you like to implement it?

Yes

Tip

_{React with 👍 to help prioritize this issue. Please use comments to provide useful context, avoiding +1 or me too, to help us triage it. Learn more here.}

Start in util/opentelemetry-util-genai/src/opentelemetry/util/genai/metrics.py at InvocationMetricsRecorder.record() and inspect how WorkflowInvocation declares and records its instrument. Add the four operation-specific instruments with the listed bucket advisories, use each metric's attribute table, and ensure agent inference_calls and tool_calls are scoped to their issuing agent. Done means tool and agent invocations emit the required metrics without relying on the client-operation histogram.