fix: gateway token double-counting with cached agents (#3306)

The cached agent accumulates session_input_tokens across messages, so
run_conversation() returns cumulative totals. But update_session() used
+= (increment), double-counting on every message after the first.

- session.py: change in-memory entry updates from += to = (direct
  assignment for cumulative values)
- hermes_state.py: add absolute=True flag to update_token_counts()
  that uses SET column = ? instead of SET column = column + ?
- session.py: pass absolute=True to the DB call

CLI path is unchanged — it passes per-API-call deltas directly to
update_token_counts() with the default absolute=False (increment).

Reported by @zaycruz in #3222. Closes #3222.

tests/gateway/test_session.py

@@ -858,6 +858,7 @@ class TestLastPromptTokens:
             billing_provider=None,
             billing_base_url=None,
             model="openai/gpt-5.4",
+            absolute=True,
         )