feat(a2a): emit aggregated turn token usage on the terminal status update

[QuentinBisson]

Fixes #2303.

What

Today per-call usage only appears as adk_usage_metadata (Go) / kagent_usage_metadata (Python) on streaming events, so every consumer wanting "tokens used" must replicate ADK's partial-event filtering and sum Gemini-shaped fields client-side.

Both executors now aggregate usage across the run's non-partial ADK events (partials are skipped; each LLM call reports its usage on its final non-partial event) and attach the total to terminal status updates under kagent_usage_total:

{"promptTokenCount": 300, "candidatesTokenCount": 50, "totalTokenCount": 350, "modelVersion": "claude-sonnet-4-5"}

The value is serialized through the exact same genai UsageMetadata mapping as the per-event usage metadata (in Go via the same toA2AMetadataMap, in Python via the same serialize_metadata_value), so consumers can share one parser. thoughtsTokenCount and cachedContentTokenCount are aggregated as well; zero counts are omitted like in the per-event shape. modelVersion is the last non-empty model version seen and is omitted when none was reported.

Coverage:

• All terminal states carry the total: completed, input-required, and failed, including failed events emitted mid-loop (Go LLM-error events, Python run-loop exceptions), so tokens burned before an error are reported too.
• Resumed tasks (HITL input-required cycles, follow-up messages on the same task) seed the accumulator from the previously persisted Task.Metadata value, making kagent_usage_total a task-lifetime total rather than a last-segment snapshot.
• Since the task update managers merge status-update metadata into Task.Metadata, the total is available both on the stream and on the persisted task (covers both variants proposed in the issue).

The key uses the kagent_ prefix (kagent-defined contract, not an ADK convention mirror); per-event adk_usage_metadata / kagent_usage_metadata are unchanged. Normalizing the field names of both keys away from the Gemini casing is deliberately left to a coordinated follow-up so the two keys never disagree on shape in the meantime.

Tests

Go: TestTurnUsageAggregatesNonPartialEvents, TestTurnUsageSkipsPartialAndEmptyEvents, TestTurnUsageShapeMatchesPerEventUsageMetadata, TestTurnUsageSeedFromTaskAccumulatesAcrossExecutions, TestTurnUsageSeedFromTaskIgnoresMissingOrMalformed.
Python: tests/unittests/test_turn_usage.py mirrors the same cases, including the shape-parity assertion against serialize_metadata_value.