Cancelled and errored rounds keep their token usage

When a conversation round was stopped mid-flight, or failed on a provider error, some providers only report their token counts in a final message that never arrived — so nothing was recorded at all, even though the prompt had already been sent and billed. Usage and cost reporting silently lost that spend.

Cancelled and errored rounds now always record at least an estimate of what was consumed — the prompt actually sent and the reply actually generated so far — clearly marked as an estimate when the provider itself never confirmed a count. Nothing is invisible to cost tracking any more.