fix(bigquery): BQAA Java P1 preview-readiness fixes (tracing, lifecycle, redaction, HITL)

[copybara-service]

fix(bigquery): BQAA Java P1 preview-readiness fixes (tracing, lifecycle, redaction, HITL)

Resolves the plugin-local P1 preview-readiness blockers in the Java BigQuery
Agent Analytics (BQAA) plugin: correctness, lifecycle, privacy-boundary, and
duplicate-telemetry defects fixable inside the agentanalytics package.
(Terminal agent/run failure analytics is out of scope; it needs new framework
plugin callbacks and will land separately.)

• Branch-safe trace state: TraceManager keeps span records in per-branch
  stacks keyed by InvocationContext.branch(), so concurrent ParallelAgent
  branches can no longer pop each other's spans. Pops are kind-checked; tool
  spans carry an operation identity and push-time parent (an event's function
  calls run concurrently); parent resolution walks the branch ancestor chain.
• No plugin-owned OpenTelemetry spans: records are ID-only (16-hex local IDs).
  Ambient OTel context is still read for trace_id/root span_id so rows stay
  joinable to Cloud Trace, but a host with an SDK exporter receives zero
  BQAA-owned spans.
• Sensitive-key redaction at the final attributes boundary: logEvent runs the
  fully assembled attributes map through the redaction walk before
  serialization (state_delta, custom tags, labels, nested extras). Redaction
  only -- no length truncation, does not set is_truncated.
• Per-invocation processor/task lifecycle: BatchProcessor.start() stores its
  ScheduledFuture; an idempotent close() cancels it so completed
  invocations no longer leak periodic tasks retaining closed processors.
• Bounded append/drain/timeout: each Storage Write RPC is bounded by
  shutdownTimeout (future cancelled on timeout); the final drain in close()
  is deadline-bounded; abandoned and after-close rows are counted
  (shutdown_timeout, after_close).
• Writer-construction and late-continuation accounting: PluginState.appendRow
  gates every append on the invocation lifecycle -- a late continuation
  completing after finalization is dropped and counted (late_after_finalize).
  StreamWriter construction failures are logged SEVERE and counted
  (writer_create_error), leaving the mapping unpopulated so the next event
  retries; writer-close ownership is bounded by admission permits.
• HITL and long-running-tool pause/resume: a synthetic adk_request_*
  FunctionCall now emits HITL_*_REQUEST (previously mislabeled _COMPLETED);
  any long-running FunctionCall additionally emits a pairable TOOL_PAUSED row
  with pause_kind and function_call_id; resumed FunctionResponse parts route
  HITL responses to HITL_*_COMPLETED and non-HITL resumes to TOOL_COMPLETED
  carrying the pause pair keys; adds the TOOL_PAUSED view and
  pause_kind/function_call_id columns on the TOOL_COMPLETED view.

Adds regression tests for concurrent ParallelAgent overlap with out-of-order
completion, per-operation tool spans, zero BQAA spans on a recording exporter,
no retained processors after many invocations, bounded close, post-finalize
append rejection, writer-construction accounting, nested/session-state
redaction, and HITL request/pause/complete naming with pair keys.