[llvm] [LoongArch] Optimize inserting fp element bitconverted from integer (PR #147043)

[via llvm-commits]

zhaoqi5 wrote:

> Does this changes also apply to la32?

Now la32 occurs error even without this commit. Shows below:
```
insert_bitcast_v4f64:                   # @insert_bitcast_v4f64
# %bb.0:                                # %entry
	movgr2fr.w	$fa1, $a0
	movgr2frh.w	$fa1, $a1
LLVM ERROR: Attempting to emit MOVFR2GR_D instruction but the Feature_IsLA64 predicate(s) are not met
```

Seems that `vector_insert` generates a `MOVFR2GR_D` which is not legal in la32.

Maybe we should modifiy it and deal with more related to `i64` type for la32 in the future.

https://github.com/llvm/llvm-project/pull/147043