[LoongArch] Add initial MachineCombiner support (#230830)

This patch add basic MachineCombiner Pass support for LoongArch and
enable it by default, including reassociative of basic arthmetic
instructions like `add/sub/mul`, logic instrunctions like `and/or/xor`
and floating-point instructions like `fadd/fsub/fmul/fmin/fmax` and
their vector counterparts.

It's worth noticing that floating-point arthmetic instruction needs
`FmReassoc` and `FmNsz` to perform reassociation, while the `fmin/fmax`
instrunctions needs `FmNoNans` because LoongArch follows the IEEE
754-2008 standard's `minNum(x, y)` and `maxNum(x, y)`, these two
operations will return `qNaN` when one of the two operands is `sNaN`, so
this will break the association law, for example:

```
min(min(3, sNaN), 1) = 1
min(3, min(sNaN, 1)) = 3
```

So we must check this, as for the -0.0 and+0.0 issue, though IEEE
754-2008 does not have any explicit declaration. Current Loongson
hardware will always return +0.0 is bigger than -0.0.

For performance metrics, when running on SPEC CPU2006 with fast-math
disabled, we get a 2.9% runtime improvment on `400.perlbench`, around 1%
for `453.povray` and `444.namd` thanks to shorter critical path
contributed by MachineCombiner. However, the `464.h264ref` did get a
constant -0.8% regressions, which comes from a association without
considering the limited execution pipeline in LA664 core, but that could
be solved a later-on TTI patch. After enable fast-math, `470.lbm` gains
a 6% runtime enhancement, but on the same time, `435.gromacs`
encountered a 1% regression due to higher register pressure cause by
reassociation. The rest of tests are basically stay unchanged.

Finally, it also depends on the schedule model which
https://github.com/llvm/llvm-project/pull/226064 added for LoongArch, so
that should be merged first.