Avoid double evaluation of pattern matchers in Chunk.collect

[flipp5b]

fs2.Chunk#collect evaluates pattern matchers and guards of a passed partial function twice because it calls pf.isDefinedAt and then pf.apply:

// ...
foreach(o => if (pf.isDefinedAt(o)) b += pf(o))
// ...

Scala collections use pf.applyOrElse/pf.runWith (depending on scala version) instead to avoid double evaluation. This PR implements the same approach in a Chunk.

It should be noted that this may change observable behavior for those who use a partial function with side-effecting matchers/guards. But hey, side-effecting matchers/guards is considered a bad practice anyway, and I doubt that there's an fs2 user relying on that double evaluation "feature."

[flipp5b]

Well, this looks rather ugly. pf.runWith returns O => Boolean so we have to transform it to O => Unit to avoid warnings. This could be rewritten as follows (at the cost of extra lambda call per Chunk item):

foreach(pf.runWith(b += _).andThen(_ => ()))

@armanbilge, which one would you prefer?

[armanbilge]

Ugly but performant is the right choice 😅

[flipp5b]

Here is some benchmark results.

1. Master

foreach(o => if (pf.isDefinedAt(o)) b += pf(o))

Benchmark               (chunkSize)   Mode  Cnt        Score       Error  Units
ChunkBenchmark.collect           16  thrpt   25  2828318.069 ± 72556.887  ops/s
ChunkBenchmark.collect          256  thrpt   25   191183.476 ±   581.751  ops/s
ChunkBenchmark.collect         4096  thrpt   25    11461.704 ±   504.370  ops/s

2. This PR

    val f = pf.runWith(b += _)
    foreach { o => f(o); () }

Benchmark               (chunkSize)   Mode  Cnt        Score       Error  Units
ChunkBenchmark.collect           16  thrpt   25  3373807.015 ± 25730.144  ops/s
ChunkBenchmark.collect          256  thrpt   25   257101.963 ±  7552.952  ops/s
ChunkBenchmark.collect         4096  thrpt   25    15417.548 ±   333.196  ops/s

3. This PR (prettified)

foreach(pf.runWith(b += _).andThen(_ => ()))

Benchmark               (chunkSize)   Mode  Cnt        Score       Error  Units
ChunkBenchmark.collect           16  thrpt   25  3420711.048 ± 90333.134  ops/s
ChunkBenchmark.collect          256  thrpt   25   242847.544 ±  4885.911  ops/s
ChunkBenchmark.collect         4096  thrpt   25    14369.581 ±   442.678  ops/s

Option 2 (this PR) shows slightly better score than the other two options.

[armanbilge]

Thank you for the very thorough PR, this is very nice work 🙏

[flipp5b]

@armanbilge Thank you for the prompt and kind feedback :)