fix(workflow): keep session identity across single-turn node contexts

[chelsealong]

Fixes #6691

Summary

When a Workflow's root agent runs multiple single-turn LlmAgent nodes
and EventsCompactionConfig.token_threshold compaction fires mid-invocation
(not on the last node), a DatabaseSessionService-backed session breaks with:

ValueError: The session has been modified in storage since it was loaded.
Please reload the session before appending more events.

InMemorySessionService does not reproduce this, because it never checks a
storage revision marker.

Root cause

prepare_llm_agent_context() in src/google/adk/workflow/_llm_agent_wrapper.py
gave every single-turn LlmAgent node its own InvocationContext, and
additionally replaced that context's session with
ic.session.model_copy(deep=False).

Session.events and Session.state are mutable containers, so the shallow
copy shared them by reference with the parent session — but
Session.last_update_time and the private storage-revision marker used by
DatabaseSessionService for stale-write detection are scalar/private
attributes, so the copy duplicated them by value.

When such a node's request-time token-threshold compaction runs
(CompactionRequestProcessor), it appends the compaction event through that
node's own session copy, advancing the revision marker on the copy only. The
parent context's session — the one later nodes copy from, and the one the
Runner's own event-consumption loop appends to — never learns about that
advance. The next append made through that stale copy is rejected by
DatabaseSessionService.append_event as a stale write.

Fix

Drop the ic.session = ic.session.model_copy(deep=False) line. Since
events/state were already shared by reference, this copy provided no
real isolation — it only caused the marker/timestamp to diverge. Every node
in the same Workflow invocation now shares the exact same Session object,
so a revision-marker advance from one node's compaction write is visible to
every other node.

Testing plan

Added test_mid_workflow_compaction_does_not_stale_later_node_append to
tests/unittests/apps/test_compaction_runner_e2e.py. It runs a two-node
Workflow (agent1 -> agent2, both single-turn) through a real
Runner.run_async against a DatabaseSessionService (sqlite), across two
turns: a short turn (no compaction), then a long turn whose estimated prompt
token count exceeds token_threshold, triggering mid-invocation compaction
on agent1 before agent2 runs. It asserts agent2's response is produced
and persisted to storage.

• Confirmed the test fails on the unmodified source with the exact reported
  error (ValueError: The session has been modified in storage since it was loaded...), by temporarily reverting only _llm_agent_wrapper.py and
  re-running.
• Confirmed the test passes with the fix applied.

$ pytest tests/unittests/apps/test_compaction_runner_e2e.py -q
...                                                                      [100%]
3 passed, 1 warning in 3.17s

$ pytest tests/unittests/workflow tests/unittests/apps tests/unittests/sessions tests/unittests/flows/llm_flows -q
1668 passed, 11 skipped, 12 xfailed, 121 warnings in 29.72s

$ pytest tests/unittests -q
11449 passed, 87 skipped, 25 xfailed, 1 xpassed, 2964 warnings, 19 subtests passed in 346.91s

pre-commit run (isort, pyink, ruff, addlicense, ADK compliance checks) passes
on both changed files.

AI-assistance disclosure

This PR was prepared with the help of an AI coding agent (Claude Code), which
investigated the root cause, implemented the fix, and wrote/verified the
regression test. I reviewed and take responsibility for the change.