Speedup ProtoWriteType.from

[fzhinkin]

While poking into #2332 I spotted that ProtoWriteType.from implementation is pretty inefficient.

There are a few ways to transform integer into a corresponding ProtoWriteType.
Corresponding benchmarks are here: https://github.com/fzhinkin/kx-serialization-parse-proto-type/blob/main/src/commonMain/kotlin/ParseProtoTypeBenchmark.kt

On JVM (see https://jmh.morethan.io/?source=https://gist.githubusercontent.com/fzhinkin/c8b3335e6ba79ec06e9acffc85230e28/raw/e420db8bbb80cf03f772862eb6a58cd7509cf617/parsing-jvm.json), both 8-aligned array-based approaches works fine.
On Native (see https://jmh.morethan.io/?source=https://gist.githubusercontent.com/fzhinkin/c8b3335e6ba79ec06e9acffc85230e28/raw/e420db8bbb80cf03f772862eb6a58cd7509cf617/parsing-macos-arm64.json), moving masking into a function works a bit faster.

Here's comparison of results proto-related benchmarks collected with existing implementation (baseline) and the implementation from this PR (optimized): https://jmh.morethan.io/?sources=https://gist.githubusercontent.com/fzhinkin/c8b3335e6ba79ec06e9acffc85230e28/raw/e420db8bbb80cf03f772862eb6a58cd7509cf617/baseline.json,https://gist.githubusercontent.com/fzhinkin/c8b3335e6ba79ec06e9acffc85230e28/raw/e420db8bbb80cf03f772862eb6a58cd7509cf617/optimized.json

Suspiciously looking ProtoBaseline.toBytesExplicit shows bimodal results, thus the results.
For parsing related benchmarks, the proposed change slightly improves performance.

[fzhinkin]

Additionally, checked how replacing the enum with raw int values will affect the performance - the difference is negligibly small and does not justify such a change.

[pdvrieze]

From an optimization perspective it may be valid to not check for typeId<0 (and maybe have a larger array of INVALIDs) to avoid the if check (instead relying on the array check that does this too).

[fzhinkin]

The array bounds check will "give" us ArrayIndexOutOfBounds, not the ProtoWireType.INVALID :)

Anyway, the implementation was implicitly based on the idea that the only call site extracts 3 least-significant bits from an int before passing it down this methods, so when this method is inlined the compiler can get rid of the if.

[sandwwraith]

Great find!

[sandwwraith]

Although this change was added in #2768 in 2024, before that, there were direct Int constants. So it is not the possible cause of slowness reported in #2332

[fzhinkin]

So it is not the possible cause of slowness reported in #2332

It's definitely not, but it improves a performance a bit without large-scale updates. You know, ten bucks is ten bucks. ;)