blocked: use an autovectorized implementation of slide_hash for wasm32

[folkertdev]

See https://godbolt.org/z/1fj8e8PfG

On nightly, the autovectorized implementation is essentially the code we expect. On 1.85.0 the loop is unrolled but not vectorized. We should use the autovectorized version (and thus get rid of the custom intrinsics) when stable rust performs the autovectorization (this is likely enabled by llvm 20, so I'd expect 1.87.0 to work)