Manual auto trait impl can be more work for the trait solver, including higher recursion depth

I accidentally hit this in https://github.com/rust-lang/rust/pull/162804#issuecomment-5686193024.

Essentially, replacing Unique<u8> with NonNull<u8> in RawVecInner<A> meant that I had to rewrite the Send, Sync, and UnwindSafe impls for it.
I initially chose to do this on RawVec<T, A> using impl<T: AuTr, A: Allocator + AuTr> AuTr for RawVec<T, A>.
This caused slight compile time regressions, as well as a compile error due to the trait solver now exceeding the recursion limit on a complex rayon iterator, likely due to the Send supertrait bound on ParallelIterator.

The benchmark that failed to compile can be found in src/tools/rustc-perf/collector/runtime-benchmarks/css/, and the failure reproduced with cargo +b67dca92df005a4ebbe8c3b45e6cfe97925fff26 check, using kennytm/rustup-toolchain-install-master.

This is dangerous especially for std, because it risks introducing regressions when refactoring the internals of commonly used types.

Note that the old impls were auto-derived from the A and PhantomData<T> fields, because Unique<u8> implemented all the relevant traits.
Now that I've pulled the manual impls back to impl<A: Allocator + AuTr> AuTr for RawVecInner<A>, it seems to no longer be regressing.
All these impls should be equivalent (if anything, I'd expect the auto impl to be more work than the manual one).

@rustbot label WG-trait-system-refactor A-auto-traits E-needs-mcve T-types

Start by reproducing the failure in src/tools/rustc-perf/collector/runtime-benchmarks/css/ with cargo +b67dca92df005a4ebbe8c3b45e6cfe97925fff26 check. Compare the manual and auto-trait implementations for RawVecInner and RawVec, focusing on the rayon iterator recursion-limit failure and compile-time regression. Done means identifying a minimal reproduction and resolving or clearly characterizing the solver regression.