Improve stream benchmark

[talex5]

Previously the multi-domain test only had one producer and one consumer. It now tries with various combinations up to 4 producers with 4 consumers. Each producer now also submits items from many fibers at the same time. This is more like a typical use of a stream, where IO domains are handling many requests by submitting CPU-intensive jobs to a pool of batch processing domains.

It now also includes some spinner fibers, to avoid measuring domain resume time. That previously dominated the results (though it happens less anyway now that the senders run multiple fibers).

On my machine, the results are:

n_send_domains,  n_iters, capacity, ns/iter, promoted/iter
             0,  1000000,        0,  258.24,        8.2001
             0,  1000000,        1,  246.76,        7.6304
             0,  1000000,      100,   69.62,        0.6136
             1,  1000000,        0, 1603.31,       30.4432
             1,  1000000,        1, 1405.07,       27.9436
             1,  1000000,      100,  368.77,        0.5061
             2,  2000000,        0, 1109.07,       10.6798
             2,  2000000,        1, 1027.75,        9.5160
             2,  2000000,      100,  272.78,        1.1984
             4,  4000000,        0, 1269.48,        7.1143
             4,  4000000,        1, 1202.87,        6.4965
             4,  4000000,      100,  441.37,        1.0967

I'm working on a PR to improve these results at the moment.