Unfortunately `ts._prefix._duration_average` == -1 for all the root tasks we care about, so this won't work.

This reverts commit e348a7c.

When a task is root-like and the previous worker is full, we don't want to use the normal `decide_worker` logic, since that only considers as candidates workers that have the deps of the dask. Since the task only has 1-5 deps, we'd only ever consider the same 1-5 workers.

Only compute `total_nthreads` when a new worker is needed, and only compute the number of dependencies once per task group. Overloads the meaning of `_last_worker` to indicate if we've decided in the past whether a TaskGroup is root-ish or not.

This reverts commit f2da0bc.

TODO test out of priority order

This reverts commit 0d1e2380b525b0ee7b4e60d9bee62f889ed5520b.

Maybe avoiding -1 is excessive; I just wanted it to still work if we changed to a Py_ssize_t

Arguably would be better to check that `ws` is in (valid_workers or all_workers); downside is that the common `all_workers` would require an O(n) search, since it's only `dict_values`, not `dict`.