the variant of rand! for an array of Float64 is slowed down by the explicit SIMD when the array is small enough

[nsajko]

Benchmarking #57101, I notice using the implementation with explicit SIMD is slowing down rand!(::Memory{Float64}) for arrays shorter than about a thousand elements. Perhaps a cutoff would be good, to only use the SIMD-ed implementation for large-enough arrays. Will investigate more deeply later.

[giordano]

There are already two different implementations, one for less than 32 elements and one for 32 or more: #55997 (comment)

[bbrehm]

To be more precise, https://github.com/JuliaLang/julia/blame/f91436eae7265a01bff17e35887cd9b8e15c8fdc/stdlib/Random/src/XoshiroSimd.jl#L169 and https://github.com/JuliaLang/julia/blame/f91436eae7265a01bff17e35887cd9b8e15c8fdc/stdlib/Random/src/XoshiroSimd.jl#L14.

The currrent simd thesholds were determined #40546 (comment). I originally set them much higher https://github.com/JuliaLang/julia/pull/34852/files# (tuned for a crappy broadwell laptop), before vanishing for some time for personal reasons (big thanks to jeff for finishing up that work I abandoned).

I think a re-tuning is in order -- ideally one would plot performance curves for all of: apple silicon, avx-2, axv-512, rasp-pi.