Notifying nexus of live repair progress updates eats way too much CPU because of creating reqwest clients

[faithanalog]

This rack is very distressed. Some disks have gone into live-repair. That distress is being made even worse because 8% of the rack's CPU is being spent notifying nexus of live-repair status, and most of that is being spent creating Reqwest clients. this is the same class of problem as oxidecomputer/omicron#5717 which was fixed by oxidecomputer/omicron#5750, but we are having it over here too.

Fixing this requires caching the Reqwest client, but it does not require caching the nexus client. This is an important distinction because caching the nexus client would cause problems when nexus moves around. But caching the reqwest client is fine, because it's just the thing that does HTTP requests, it doesn't care where they're going.

In short, we want to be calling nexus_client::Client::new_with_client() instead of nexus_client::Client::new() here

crucible/upstairs/src/lib.rs

Lines 1816 to 1819 in da333e9

	Some(nexus_client::Client::new(
	&format!("http://{}", nexus_address),
	log.clone(),
	))

A Reqwest client is internally Arc'd, so we just need one laying around that we can clone every time we make a nexus client

@iliana tagging you to make sure I'm explaining this right

[iliana]

Yep that's right.