Currently merge_all_threads can result in out-of-memory, when concurrent is not Infinity and the source submits items faster than the mapper handles items.

I wonder if we could change the type of concurrent from usize to a struct:

pub struct ConcurrentConfig<F, Item> {
  usize max_number_of_inner_observers;
  usize max_buffer_size;
  F backpressure_handler;
}

impl<F, Item> ConcurrentConfig<F, Item>
where
    F: FnMut(Item) -> () {
}