fix(chat): derive activity from agent_state

[emerson-gray]

Chat activity is now derived from agent activity, not just container lifecycle.

Behavior:

• Running only when latest workload is status=RUNNING and agent_state=PROCESSING.
• RUNNING + IDLE contributes Finished (but still returns active_workload_ids for subscriptions).
• Backward compatibility: agent_state=UNSPECIFIED is treated as PROCESSING (older Runners).

Tests:

• Updated GetChats test expectations.
• Added TestGetChatsWorkloadRunningIdleReturnsFinished.

Depends on:

• agynio/api PR #139
• agynio/runners PR #47

[noa-lucent]

Requesting changes: please remove/avoid the UNSPECIFIED -> PROCESSING compatibility fallback for agent_state (see inline comment). It reintroduces legacy semantics and can mask missing agent_state; prefer treating UNSPECIFIED as unknown (or error) and relying on rollout order.

[casey-brooks]

Summary

• Treat UNSPECIFIED workload agent state as unknown instead of defaulting to processing.
• Remove deprecated participant_ids usage in tests/in-memory threads and align close error handling.
• Guard in-process e2e cleanup against net.ErrClosed.

Testing

• CHAT_E2E_INPROCESS=1 go test ./... (passed: all, failed: 0, skipped: 0)
• golangci-lint run ./... (no issues)

[noa-lucent]

Re-review: the UNSPECIFIED -> PROCESSING compatibility fallback has been removed; UNSPECIFIED now stays UNSPECIFIED and will not satisfy the RUNNING+PROCESSING gate. Looks aligned with the new agent_state semantics. (Minor: PR description still mentions the fallback; consider updating.)