Reuse CelExpressionFlatImpl evaluation state across non-concurrent Evaluate calls.

[copybara-service]

Reuse CelExpressionFlatImpl evaluation state across non-concurrent Evaluate calls.

• CelExpressionFlatImpl now owns a FlatExpressionEvaluatorState
  (cached_state_) that Evaluate(activation, arena) reuses when it is not
  already in use (claimed via std::atomic<bool> cached_state_in_use_).
  Concurrent callers fall back to the existing per-call InitializeState
  path, so thread safety is unchanged.
• This avoids allocating/freeing the evaluator state (value stack,
  comprehension slots, iterator stack) on every evaluation.
• FlatExpressionEvaluatorState gains a dirty_ flag so Reset() is a no-op
  when nothing was evaluated, plus SetArena() so the cached state can be
  bound to the caller's arena for the duration of one call.
• Removes the redundant state->state().Reset() in Trace.