Dynamic adaption of average task duration

We currently experience an issue with very long-running tasks, when using `adaptive_target` to figure out how many workers the dask graph needs.
The estimation by `adaptive_target` is currently wrong as long as no task of the TaskPrefix has been executed yet.
Could we change the estimation to correct dynamically and propagate a new average across the cluster as soon as the execution of one task e.g. already takes twice the estimated time?