Better lowering for shuffleFloatX4# and shuffleDoubleX2#
<!--
READ THIS FIRST: If the feature you are proposing changes the language that GHC accepts
or adds any warnings to `-Wall`, it should be written as a [GHC Proposal](https://github.com/ghc-proposals/ghc-proposals/).
Other features, appropriate for a GitLab feature request, include GHC API/plugin
innovations, new low-impact compiler flags, or other similar additions to GHC.
-->
## Motivation
[Current implementation](https://gitlab.haskell.org/ghc/ghc/-/blob/bfa6b70f27dc2ce7fc890ec71103c40f66497c77/compiler/GHC/CmmToAsm/X86/CodeGen.hs?page=3#L2043-2109) of `shuffleFloatX4#` and `shuffleDoubleX2#`
* doesn't make use of specialized instructions like `(V)UNPCK{L,H}{PS,PD}` (they have shorter encoding than `(V)SHUF{PS,PD}`), and
* require `-mavx`. The dependence on AVX is non-essential.
They should use specialized instructions, and be implemented on SSE2 (our baseline ISA for x86).
## Proposal
Re-implement `shuffleFloatX4#` and `shuffleDoubleX2#` to satisfy the criteria above.
issue