Add virtual disk timeseries schema

[bnaecker]

This adds a new set of timeseries that track block operations on virtual disks. This builds on and replaces the pre-existing Crucible data, adding more information about the disk and instance it's attached to. It also tracks I/O latencies and sizes in histograms.

[bnaecker]

I've used this while testing oxidecomputer/propolis#746, which has examples of the data this publishes. But to summarize, we have

• Cumulative counters for the number of reads, writes and flushes
• Cumulative counters for the number of bytes read and written
• Cumulative counters for the number of failed reads, writes and flushes. These are additionally broken out by a string indicating the failure reason.
• Histogram of I/O latencies for reads, writes and flushes. The bins are actually defined in Publish richer virtual disk statistics to oximeter propolis#746, but they are log-linear. There are 10 bins in each power of 10, from 1 microsecond up to 10 seconds.
• Histogram of I/O sizes for reads and writes. Again the bins are in Propolis, but there are straight power-of-two bins from 4KiB to 1GiB. The is an additional bin on either side capturing anything outside that range.

The thing I'd like some input on is related to #5267. I had originally put the sled identifiers on this timeseries. Adam raised some concerns about actionability of those fields; how they might confuse developers, who don't really know about and can't see the sleds; and that it will complicate the rough idea for an authz model. I agree with those points, but would appreicate some other perspectives too.

[bnaecker]

It occurs to me we should reduce the lowest I/O size histogram bin to 512 bytes. That's the minimum block size the control plane allows, so should be the smallest actual block operation size.

[bnaecker]

Holding off on this while we sort out how to cut the R10 release branch / commit with the new process.

[bnaecker]

Looks like we ran afoul #6300 of here. Going to rerun that workflow.