fix(telemetry): propagate OTel trace context so all spans share one trace

[fede-kamel]

Closes #244. Draft until Langfuse repro lands this afternoon —
we'll cross-check it against this fix + add it as a second regression
test before marking ready-for-review.

Summary

TelemetryHook was calling tracer.start_span(...) for invocation,
iteration, and tool spans without ever attaching them to OTel's current
context. Every span emerged as its own trace root, so Langfuse / Jaeger /
Tempo rendered N separate traces per agent run instead of one tree.

Reported by @Luigis while wiring up Langfuse.

Fix

opentelemetry.context.attach(trace.set_span_in_context(span)) after
starting each parent span, paired with context.detach(token) on end.
This pushes the span onto OTel's context stack so subsequent start_span
calls within the same async scope automatically parent to it.

Hook	Before	After
on_before_invocation	starts span, doesn't attach	starts span + attaches to OTel context
on_after_invocation	ends span	ends span + detaches context
on_iteration_start	starts span, doesn't attach	starts span + attaches (auto-child of invocation)
on_iteration_end	ends span	ends span + detaches
on_before_tool_call	starts span at root	starts span (auto-child of iteration)
on_after_tool_call	ends span	unchanged — tool span needs no context bookkeeping

Regression tests

Three new tests under TestTraceContextPropagation using OTel's
InMemorySpanExporter:

1. test_all_spans_share_one_trace_id — asserts the 3-span run
   (invocation + iteration + tool) shares exactly one trace_id.
2. test_tool_span_is_child_of_iteration_span — verifies the parent
   chain via span.parent.span_id: invocation has no parent (root),
   iteration's parent = invocation, tool's parent = iteration.
3. test_invocation_context_detaches_after_run — confirms the context
   is detached on completion so the caller's downstream spans aren't
   parented to ours.

Tests gracefully skip when opentelemetry-sdk isn't installed in the
test env (the hook only needs opentelemetry-api at runtime).

Test plan

• hatch run check — ruff format + ruff check + mypy clean
• 31 / 31 telemetry tests pass (including 3 new regressions)
• 4788 full unit suite green
• Cross-check against Luigi's Langfuse repro when it lands

Out of scope (filed for later)

The hook still doesn't currently set the service.name resource (just
attaches it as a span attribute). Spec-wise it should live on the
resource, not the span — but moving it requires a Resource on the
TracerProvider, which is the caller's concern. Leaving as-is.

[luigisaetta]

OK, I have done the tests @fede-kamel the issue seems to be solved!