We are doing direct transmutes between `OsStr` and `[u8]`.
https://github.com/rust-lang/rust/pull/95290 would make this natively
supported but I got tired of waitin for it.
This only saves about 1/4s off of `cargo build`.
This took 2.9 KiB off of `cargo bloat --release --example git`
For now, we are still treating `clap` as the user facing API for both
builder and derive, making this an internal change as we don't expect
this to negatively impact builder build times all that much. We can
re-evaluate at a later time and consider having distinct top-level
crates for builder and derive.
Looking at `--timings` on my machine
- `clap` only took 0.04s to build and it happened in
parallel to `clap_builder` codegen
- this saved 1.7s for derive build times, with `clap_builder` building
in parallel to `syn` and `clap_builder` and `clap_derive` finishing
around the same time.
This was discussed some at https://rust-lang.zulipchat.com/#narrow/stream/220302-wg-cli/topic/clap.20build.20times.20and.20.60clap_derive.60.3A.20a.20crazy.20idea
