This patch fixes the flaky test variables-locations.test. The test began flaking after the fix of the PR38857 issue. It have happened because in PDBLocationToDWARFExpression we treat a VFRAME register as a LLDB_REGNUM_GENERIC_FP, which leads to EBP on Windows x86, but in this case we can't read locals relative to EBP. Moreover, it seems that we have no alternative base, so I just have changed double with float to avoid alignment.
By the way, what do you think, how can we make LLDB support aligned stacks? As far as I know, similar alignment problems are reproducible on non-Windows too.