Done to avoid copying potentially hige inputs and functions to
the processes when all they need to do is crunch some silly small
numbers. HUGE performance gain while reducing memory usage a lot
for benchmarks with bigger data in use.

Also use Task.async_stream instead of our good old `Parallel.map`
as with the dawn of `inputs` I have seen people do 30+ scenarios
(10 benchmarks with 3 inputs will do that to you), where if we're
running on a 4 core system we might be doing too much in parallel
also potentially skyrocketing memory consumption.

This includes a slight workaround to preserve the names of inputs
as people may be relying on those. Doesn't feel great but...
somehow reasonable?

Now, that scenario is finally fast/fastest in that lil benchmark:

```
Name                        ips        average  deviation         median         99th %
Formatter.output         248.71        4.02 ms     ±4.58%        4.03 ms        4.59 ms
sequential_output        248.38        4.03 ms     ±3.16%        4.00 ms        4.53 ms
format & write           247.77        4.04 ms     ±6.76%        4.02 ms        4.35 ms

Comparison:
Formatter.output         248.71
sequential_output        248.38 - 1.00x slower +0.00539 ms
format & write           247.77 - 1.00x slower +0.0153 ms
```