refine CDN invalidations after build

[jsha]

Following up on #1825:

I notice there are a few ? operators that come after the build happens, but before the CDN invalidation. For instance, purging the build dir and the local cache:

https://github.com/rust-lang/docs.rs/pull/1825/files#diff-b06b44e6583254ce69af8cedabc04a73a97599d593676e1bbad7ee442240ccf1R502-R506

In the unlikely event one of these fails, we would fail to invalidate. Probably the CDN invalidation should happen regardless of those errors.

Relatedly: Right now it looks like we'll do a CDN invalidation even for a build that failed. But a failed build won't affect any page except /crate/<name>/<version>/builds. Depending on how frequent failed builds are, we could save some money on invalidations by making /builds max-age=0, and not invalidating cache on builds.

[Nemo157]

Failed builds affect the version list in the crate dropdown (at least without #1772).

Also the version list on the left of the /crate/<name>/<version> page, which probably doesn't make sense to dynamically load.

[syphar]

invaliding when we use ? is definitely a gap in the current logic, I would fix this after we merged #1856,
but before we introduce the default TTL ( = before we actually start caching)

[syphar]

@jsha perhaps I'm missing something because of the early morning, but:
is this here still an issue? The diff-link above doesn't show me any example I need, perhaps it became invalid because of some force-pushes.

I'm mostly talking about the ? piece where we would not create the needed invalidation in certain error situations.

[syphar]

correction: I just see I might have solved it myself in #1864 , so this doesn't block increasing the TTL any more.