[llvm] [RISCV] Enable (non trivial) remat for most scalar instructions (PR #162311)

[Alex Bradbury via llvm-commits]

asb wrote:

Here are dyn instcount diffs for an rva22 build of SPEC 2017:

```
Benchmark                  Baseline       This PR   Diff (%)
============================================================
500.perlbench_r         179038504795    179015518664     -0.01%
502.gcc_r               221238744000    221171512681     -0.03%
505.mcf_r               134655886612    137150096559      1.85%
508.namd_r              217623220031    217709012429      0.04%
510.parest_r            291729214114    291727407077     -0.00%
511.povray_r             30983012423     30981981254     -0.00%
519.lbm_r                91217999797     90475828912     -0.81%
520.omnetpp_r           137704191763    137702616618     -0.00%
523.xalancbmk_r         284738544130    284738525706     -0.00%
525.x264_r              379871669079    379399368988     -0.12%
526.blender_r           659313110004    659133873329     -0.03%
531.deepsjeng_r         349454510283    349291680562     -0.05%
538.imagick_r           238568576282    238568485759     -0.00%
541.leela_r             405707905587    405700609695     -0.00%
544.nab_r               398215408162    398165080506     -0.01%
557.xz_r                129537393796    129925509975      0.30%
```

Looking at the static assembly diff it is large due to lots of very tiny regalloc changes. The obvious outlier is mcf, which I'll need to report back on after having a closer look.

https://github.com/llvm/llvm-project/pull/162311