Speeding up Scala 3 Writing macros

[lihaoyi]

Fixes #389 and #388

This PR moves a bunch of logic from runtime to compile time, and optimizes whatever runtime logic remains. It also consolidates the logic with the Scala 2 logic as much as possible, e.g. re-using writeSnippet to manage writes and CaseObjectContext/HugeCaseObjectContext to manage validation and error reporting during a read

Running some ad-hoc benchmarks on my laptop, ./mill bench.jvm.run, it gives a significant speedup on reads and writes, and brings it close to the Scala 2.13 numbers (higher is better):

Benchmark (5000ms)	Scala 3 Before	Scala 3 After	Scala 2
upickleDefault Read	637	1065	1403
upickleDefault Write	839	1452	1549
upickleDefaultByteArray Read	582	1172	1126
upickleDefaultByteArray Write	847	1218	1277
upickleDefaultBinary Read	925	3853	3844
upickleDefaultBinary Write	1252	3117	3666
upickleDefaultCached Read	620	1300	1412
upickleDefaultCached Write	829	1555	1588
upickleDefaultByteArrayCached Read	575	1182	1095
upickleDefaultByteArrayCached Write	838	1223	1297
upickleDefaultBinaryCached Read	928	3825	3885
upickleDefaultBinaryCached Write	1266	2907	3674

Note that the generated code, especially for reads, is still not as optimized as the Scala 2 versions:

1. I couldn't figure out how to generate an anonymous class with typed fields in Scala 3, so I'm putting things in an Array[Any]
2. I couldn't figure out how to generate match statements, so I generated a Map[K, V] and look it up at runtime.

Fixing this issues and moving the reading logic into the macro should also be possible, but can happen in a separate PR

All existing tests pass. Added a regression test for the recursive Scala 3 scenario that hangs on master. Also moved a bunch of AdvancedTests into the shared folder now that they work

[lihaoyi]

Had to make these un-typed because I couldn't figure out how to make the Scala 3 macro quotes/splices play nicely with the path dependent prefix.Writer type. A bit gross but shouldn't affect perf or correctness

[lihaoyi]

I couldn't figure out how to subclass the path dependent prefix.CaseW type in the macro, so I had to instantiate it in the inline part first and then call into the macro after. That means the call to writeSnippet is no longer from the subclass, and it thus can no longer be protected

[bishabosha]

this won't work for ADT enums, the solution here is to assert all cases are singleton:

summonAll[Tuple.Map[m.MirroredElemTypes, [S] =>> S <:< Singleton]] // assert all singleton

maybe there is a more erased way to do this

[lihaoyi]

I'll leave this over the weekend, and if there aren't any more comments I'll merge it then. All the existing/new/newly-enabled tests pass, and the benchmarks demonstrate the non-functional improvements, so I can't have broken anything too badly

The issue noticed by @bishabosha is a limitation in existing code, which is why it was not picked up by any tests (since it never worked). Fixing that can come in a follow up

[lefou]

This looks good to me, although I'm myself not experienced in Scala 3 Macros. But tests pass and benchmark looks promising.

After we merge this, I'd like to cut a next milestone release 3.0.0-M2 to get this changed version tested in as many places as possible.

[lefou]

There seem to be issues with this change. E.g. in Ammonite, the integration tests don't succeed.

• Update upickle to 3.0.0-M2 Ammonite#1320