Track input/output token counts

[simonw]

With new LLM feature:

• Abstract out token usage numbers llm#610

[simonw]

I got stuck writing a full test for this, because there's something weird about the current pytest-recording setup - look at this:

llm-gemini/tests/test_gemini.py

Lines 15 to 23 in cfe61b5

	assert response.response_json == [
	{
	"candidates": [
	{"content": {"parts": [{"text": "Percy"}], "role": "model"}}
	],
	"usageMetadata": {"promptTokenCount": 10, "totalTokenCount": 10},
	"modelVersion": "gemini-1.5-flash-002",
	}
	]

Why does that have input tokens but no output tokens?

I couldn't figure that out, so I'm going to assert for input tokens only for the moment.