Build a fleet compute and token cost model

The model is four small functions and a fleet file. Every rate is converted to USD per second (or per token) as it is loaded, and every function takes seconds; hours, GB-hours and "per million" exist only in rates.toml. The decision that matters most is modelling Fluid's two meters separately: CPU is billed only while code runs, memory for the whole wall time, and in-function concurrency divides only the memory bill.