nativeGen/AArch64: Fix sign extension in MulMayOflo

Previously the 32-bit implementations of MulMayOflo would use the
a non-sensical sign-extension mode. Rewrite these to reflect what gcc 11
produces. Also similarly rework the 16- and 8-bit cases.

This now passes the MulMayOflo tests in ghc/test-primops> in all four
widths, including the precision tests.

Fixes #23721.