fix(sysfs): accept delayed mdraid sync_completed state

[happysnaker]

Summary

• treat sync_completed=delayed as a valid mdraid state instead of failing the entire Mdraids() parse
• keep numeric progress parsing unchanged for the normal a / b form
• add a focused regression test that builds a minimal mdraid sysfs tree with sync_completed=delayed

Why

When multiple md arrays share physical devices, Linux can expose sync_completed as delayed while a check or resync is deferred. The current parser only accepts none or a / b, so one delayed array causes sysfs.Mdraids() to fail outright.

This change treats delayed like the existing non-progress sentinel none: there is no numeric completion fraction to parse yet, but callers still get the rest of the array metadata.

Closes #770

[happysnaker]

Maintainer context: Linux can expose sync_completed as the literal delayed when a check/resync is deferred because another md array sharing the same devices is already syncing. In that state there is no numeric progress fraction to parse yet, but the rest of the array metadata is still valid. This patch therefore treats delayed the same way the parser already treats none: keep returning the mdraid entry instead of failing Mdraids() for the whole scrape.