fix(llm-trace): stash litellm tool-call usage so token cost survives arg-parse failures

[r266-tech]

Completes the tool-calling path for #2396 (closes #2387 — keep the provider-billed token cost on error traces).

#2396 made LiteLLMLLM.call() stash the response usage right after the billed completion so that a local failure (e.g. a length check or JSON parse) doesn't drop the tokens the provider already charged for. call_with_tools() was left without that stash:

• the response is awaited and billed at the top of the try,
• then tool-call arguments are parsed with an unguarded arguments = json.loads(arguments) (litellm_llm.py:423),
• a malformed arguments string raises json.JSONDecodeError after billing; the broad except Exception classifies it non-retryable and re-raises, and the wrapper error path reads current_response_usage() → None → records input/output_tokens = 0/0.

That is exactly the accounting loss #2387 set out to eliminate, just on the tool-calling route — and LiteLLMRouterLLM inherits it (it overrides _acompletion/_build_common_kwargs/_stage_label/_resolve_completion_model, not call_with_tools).

This adds the same stash_response_usage(_usage_from_litellm_response(response)) immediately after the billed response, mirroring call() and the anthropic/gemini call_with_tools paths. On the success path the usage is recomputed downstream as before; on the error path the provider cost now survives so the total-vs-effective-vs-failed token split #2387 asked for stays accurate.