[llvm] [GlobalISel] emit G_BITCAST in widenScalarUnmergeValues when SrcTy is float (PR #204541)

[Kacper Doga via llvm-commits]

================
@@ -4400,4 +4400,37 @@ TEST_F(AArch64GISelMITest, LowerFMinimumExtLLT) {
   LLT::setUseExtended(false);
 }
 
+TEST_F(AArch64GISelMITest, WidenScalarUnmergeValuesFloat) {
----------------
varev-dev wrote:

Now that AMDGPU has switched to extended LLT and exposes the widenScalar rule on `G_UNMERGE_VALUES` I would replace the unit test with MIR/IR tests.

https://github.com/llvm/llvm-project/pull/204541