when TensorRT-LLM clearly wins

[rasterbench]

I'd approach it in a boring order. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at continuous batching. Turns out 'just one more test' was a lie.