CodeGen: Fix double counting bundles in inst size verification (#191460)
The AMDGPU implementation handles bundles by summing the
member instructions. This was starting with the size of the
bundle instruction, then re-adding all of the same instructions.
This loop is over the iterator, not instr_iterator, so it should
not be looking through the bundled instructions. Most of the other
uses of getInstSizeInBytes are also on the iterator, not the
instr_iterator so the convention seems to be targets need to handle
BUNDLE correctly themselves.