aarch64: use portable SIMD intrinsics where possible

We have a number of portable SIMD intrinsics in core::intrinsics::simd. The advantage of using those is that they have to only be implemented once by non-LLVM backends (Miri, cranelift, gcc) and then they work consistently across all architectures. The x86 parts of stdarch thankfully make good use of them, which heavily reduces the number of intrinsics Miri and the other backends have to special-case by hand.

It would be good to do the same for aarch64. AFAIK aarch64 has completely standard float semantics including the behavior on subnormals, so there shouldn't be a reason to use architecture-specific intrinsics. (arm-32 is a different case since the neon intrinsics there flush subnormals to 0, which makes it technically not correct to use the portable intrinsics. But also at least LLVM doesn't implement the portable intrinsics properly on arm-32 so they actually do flush subnormals, except if optimizations kick in in which case we can get unsoundness.)

Start by reviewing core::intrinsics::simd and the existing x86 implementations in stdarch, then compare how the aarch64 intrinsics are currently provided. Identify operations whose aarch64 floating-point semantics match the portable intrinsics, and consider the arm-32 subnormal behavior separately. Done means the applicable aarch64 operations consistently use portable SIMD intrinsics without changing their documented semantics.