avoid_web_libraries_in_flutter perf improvements: fix pubspec state caching

[pq]

Context:

Before this change, we were parsing pubspecs for each and every compilation unit because our caching mechanism was faulty. This fixes that to limit parsing ideally to once per context root.

For reference, a typical benchmark run before this change was ~250ms. Because the code being benchmarked is not in a flutter package, with this change we're seeing no count. (See notes below.)

Notes:

• benchmarks are currently not capturing time in registerNodeProcessors(...) so this lint is not measured by active benchmarks
  • this is fixed w/ https://dart-review.googlesource.com/c/sdk/+/199372 and will get picked up in the next analyzer push
  • benchmarking in general should be improved to include more sources (see Add a big codebase to the benchmarks. dart-lang/sdk#57576)
• when benchmarking flutter/packages/flutter_web_plugins (which is a flutter package and so is benchmarked using the current mechanism) we see an improvement from 20ms to ~0.
• since this lint is actively in the flutter list, this improvement will affect LOTS of users

TL;DR: I think this approach is sound and a major improvement but benchmarking is lacking. It will be improved on the next analyzer push but we'll need to update what we're benchmarking against to get better results moving forward.

/cc @bwilkerson

/fyi @goderbauer

[coveralls]

Coverage increased (+0.003%) to 93.898% when pulling 6bec20b on avoid_web_lib_check_fix into b448610 on master.

[bwilkerson]

I agree, the changes look good.

We probably need to benchmark different rules against different code bases, given that some rules only apply to certain kinds of code bases.

[pq]

We probably need to benchmark different rules against different code bases

Agreed. We've got a tracking issue for benchmarking in dart-lang/sdk#57576. It'd be great to bump this up in the coming quarters.