perf(serde_v8): build with codegen-units=1

[littledivy]

Lets LLVM optimize code better. Decreases ~160ns from ser_struct_v8_manual bench. Build time impact is negligible.

[AaronO]

We previously built everything in release with codegen-units=1, #10039 removed that but #10086 reverts it, which means that this proposed optimization is actually a temporary regression that should disappear.

[AaronO]

Also ser_struct_v8_manual isn't a benchmark we want to improve per se, it's a baseline/reference bench that compares naively/manually serializing a rust struct to v8 "by hand".

Since serde_v8 aims to be a "zero overhead" abstraction, it should be competitive or even beat manual serialization code (thanks to internalized keys, Key Cache, etc...).

[piscisaureus]

I guess this is superseded by #10086.