feat: add ADK 1.19/2.0 cross-version regression suite and lazy-import compat fixes

Adds a 44-test regression suite targeting every veadk↔ADK seam where 1.19 and
2.0 differ, and fixes two compat bugs the suite caught when run against ADK
2.0.

New tests (tests/test_adk_compat_regression.py)
-----------------------------------------------
- A. Direct coverage of every public helper in veadk.utils.adk_compat that
  test_adk_compat.py didn't already exercise (get_adk_version, is_adk_gte,
  should_use_async_db_drivers, llm_request_has_field, plus the getter path
  for get_event_function_responses).
- B. Edge cases for the event extractors: parts=None / empty / mixed, no
  content, broken getters falling back to part traversal.
- C. Integration against real ADK Event / Content / Part objects so we
  notice if ADK silently changes its public surface.
- D. Agent.run override gated by is_adk_gte("2.0.0") (skipped on v2 by
  design; verifies NotImplementedError still surfaces on v1).
- E. ArkLlm version-branching: ImportError when LlmRequest lacks
  previous_interaction_id; get_previous_interaction_id on real LlmRequest.
- F. tool_attributes_extractors fallback variants (model_dump path, empty
  sentinel, attribute object, missing attrs).
- G. Runner.intercept_new_message integration with a synthetic LLM:
  session_service is InMemorySessionService for local STM, create_session
  contract, end-to-end yield, None-event filtering, part.text=None
  tolerance.
- H. ADK public-surface assumptions (version module, LlmRequest.model_fields,
  Event.get_function_calls/responses).

Compat fixes caught by running the suite under ADK 2.0
------------------------------------------------------
- veadk/agent.py: gate the legacy Agent.run NotImplementedError override on
  `not is_adk_gte("2.0.0")`. ADK 2.0 promotes BaseAgent.run to a @Final
  async generator that the Workflow/NodeRunner engine invokes internally;
  overriding it breaks workflow execution.
- veadk/utils/patches.py: walk `mod.__dict__` instead of `dir(mod)` +
  `getattr` when patching tracers across google.adk.* modules. ADK 2.0's
  `google.adk.tools` package defines `__getattr__` for lazy submodule
  loading; the old dir+getattr pattern triggered every lazy module, including
  optional ones like discovery_engine_search_tool that need google-cloud-*
  deps veadk doesn't ship.
- veadk/evaluation/eval_set_recorder.py: defer the
  `from google.adk.cli.utils import evals` import to the call site. ADK 2.0
  has that module top-import gcs_eval_set_results_manager unconditionally,
  which requires google-cloud-storage — needlessly tainting any veadk caller
  who only imports veadk.Runner.

Regression results
------------------
Both versions show identical counts; each skips exactly the one test that
doesn't apply to its API surface (the suite is symmetric across versions):

* ADK 1.19.0: 195 passed + 1 skipped
  (skip: test_get_previous_interaction_id_with_real_llm_request —
   v1 LlmRequest lacks previous_interaction_id)
* ADK 2.0.0:  195 passed + 1 skipped
  (skip: test_agent_run_raises_notimplemented_on_legacy_adk —
   override is removed on v2 by design)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: tiantt