[antonrh]

A regression from the release fix earlier this week, found by a second review
pass over it.

release() ran the finaliser before dropping the instance, so a finaliser
that raised left the instance cached with its resource already closed. The
next resolution handed that closed instance out:

release raised: finaliser blew up
still cached: True
events after resolving again: ['open']     # no second open

The instance is dropped first now, and the same run reads:

still cached: False
events after resolving again: ['open', 'open']

The ordering also aborted reset() halfway, leaving later singletons neither
closed nor reset, with nothing to say which half had run. reset() now
releases them all and raises the first failure afterwards:

events: ['a open','b open','c open','a close','b close','c close']
still cached: none

Before this week's change reset() used pop, which cannot raise, so it
always finished. This restores that guarantee without giving up the closing.

[codspeed]

Merging this PR will not alter performance

✅ 17 untouched benchmarks

---

_{Comparing fix-release-drops-instance (60bf1d0) with main (cd2a974)}