Start byte-based backpressure earlier

[mkeeter]

This PR implements a suggestion from RFD 445 of having byte-based backpressure start earlier (see the section titled Whither latency?).

It also caps the max byte-based backpressure, so that we'd timeout sooner if a Downstairs goes away during a large-write workload (see Problematic Bytes-In-Flight Delay)

Previously, the system stabilized at about 1.8 GiB of in-flight data. With the tuning in this PR, it stabilizes at around 500-600 MiB instead. This means that when we switch from writes to reads, we doesn't need to wait as long!

I see no performance impacts. For reference, here's a quick fio test run:

1M WRITE: bw=723MiB/s (758MB/s), 723MiB/s-723MiB/s (758MB/s-758MB/s), io=42.5GiB (45.6GB), run=60099-60099msec
4K WRITE: bw=28.3MiB/s (29.7MB/s), 28.3MiB/s-28.3MiB/s (29.7MB/s-29.7MB/s), io=1698MiB (1781MB), run=60009-60009msec
1M WRITE: bw=772MiB/s (810MB/s), 772MiB/s-772MiB/s (810MB/s-810MB/s), io=45.3GiB (48.6GB), run=60025-60025msec
4M WRITE: bw=754MiB/s (791MB/s), 754MiB/s-754MiB/s (791MB/s-791MB/s), io=44.3GiB (47.5GB), run=60107-60107msec
4K READ: bw=27.5MiB/s (28.9MB/s), 27.5MiB/s-27.5MiB/s (28.9MB/s-28.9MB/s), io=1653MiB (1733MB), run=60003-60003msec
1M READ: bw=629MiB/s (659MB/s), 629MiB/s-629MiB/s (659MB/s-659MB/s), io=36.9GiB (39.6GB), run=60036-60036msec
4M READ: bw=737MiB/s (773MB/s), 737MiB/s-737MiB/s (773MB/s-773MB/s), io=43.3GiB (46.5GB), run=60130-60130msec

This is roughly the same as numbers from main here.

(#1167 remains unaddressed by these changes)