don't use read_unaligned to read from an array casted to *const Simd

[programmerjake]

We should create a simd_read_array intrinsic because for sizeof(Simd<T, N>) > sizeof([T; N]) (which can happen until #319 is fixed) read_unaligned is probably UB due to being able to read the bytes beyond the end of the input array -- the padding in the Simd<T, N>.

We need an intrinsic rather than just using memcpy because the intrinsic will generate llvm's load instruction with vector type (llvm guarantees vector load won't read padding if the load's align is small enough), whereas memcpy may end up using less efficient array-typed loads which sometimes use scalar code.

https://rust-lang.zulipchat.com/#narrow/stream/257879-project-portable-simd/topic/splat.20no.20longer.20compiles.20for.20release.20builds/near/352101044

[scottmcm]

I think that with rust-lang/rust#111999 it might be fine to do this as an array copy? This:

const N: usize = 16;
pub fn simd_from_array(a: [u32; N]) -> Simd<u32, N> {
    let mut v = Simd::default();
    *v.as_mut_array() = a;
    v
}

Generates a vector load/store now: https://rust.godbolt.org/z/hYv4T17TP

define void @_ZN7example15simd_from_array17h5fc2848bbb0cc1d0E(ptr noalias nocapture noundef writeonly sret(<16 x i32>) align 64 dereferenceable(64) %_0, ptr noalias nocapture noundef readonly align 4 dereferenceable(64) %a) unnamed_addr #0 {
  %v.0.copyload = load <16 x i32>, ptr %a, align 4
  store <16 x i32> %v.0.copyload, ptr %_0, align 64
  ret void
}

(Though it'll fallback to a safe memcpy for non-pot sizes because rust-lang/rust#115236 found that non-pot vector loads & stores go very poorly on many platforms.)