Resolve paths for views asynchronously

[aredridel]

This eliminates the only post-startup synchronous I/O in Express.

In addition, this makes the view.lookup method non-private, intended to be used
by view engines to resolve paths to partials, and, if the View class is
extended, could be used for other resources such as internationalized messages.

[dougwilson]

Nice! Anyway we can add to this a way to prevent stampeding (which the sync version protected from) to the default lookup implementation?

[aredridel]

Oh, yes! I think I could work that in.

[aredridel]

Updated.

[gabrielcsapo]

👍

[bjohansebas]

Hey, @aredridel could you rebase this PR, please?

[bjohansebas]

@expressjs/express-collaborators @wesleytodd @UlisesGascon Any objections to releasing this in the next minor version of Express?

[jonchurch]

Does someone want to own last mile on this? Its in my list of stuck PRs, so either working to understand and land, or understand and close
cc @expressjs/express-collaborators

Im unassigning @blakeembrey bc its been so long idk why I assigned it to him and I doubt he does either

the (current) test fails are unrelated to the PR

[bjohansebas]

It’s been a long time since the last time I reviewed it, I’d like to take another look in case I missed something. I’ll put it on my TODO list.

[bjohansebas]

A few things I changed were:

• Removed lookupMain and restored app.render to its simpler form.
  The original PR introduced a View.prototype.lookupMain(cb) method that app.render had to call explicitly, with an if (!view.path) { lookupMain(...) } branch. I moved that logic into View.prototype.render instead (lazy lookup on the first render). This lets app.render go back to not knowing anything about how views are resolved, reduces the internal API surface, and means custom View implementations don't need to implement a new method.

• Treat all stat errors as cache misses.
  The original implementation aborted rendering for any stat error other than ENOENT (for example, ENOTDIR if one of the view roots points to a file). I restored the synchronous behavior, where any stat error is treated as "not found" and the search simply continues.

• Preserved the custom View contract and error behavior.
  Since the original PR always called lookupMain, it changed the contract for app.set('view', ...). I kept it identical to the synchronous implementation, so existing custom View classes continue to work unchanged.

On top of that, I added a couple of improvements beyond the original implementation:

• Coalesced concurrent renders.
  The original implementation didn't do this. If N concurrent renders requested the same cold view, it would perform N independent lookups. I collapsed them into a single lookup, which can reduce I/O by up to 50× during bursts.

• Handled invalid view roots without introducing new failures.
  The original implementation didn't handle non-string entries in the views array. I restored the same TypeError behavior as the synchronous implementation instead of introducing a different failure mode.

Overall, the goal of these changes was to avoid introducing breaking changes while still improving the implementation.