locking/atomic: cmpxchg: Make __generic_cmpxchg_local compare against zero-extended 'old' value

__generic_cmpxchg_local takes unsigned long old/new arguments which
might end up being up-cast from smaller signed types (which will
sign-extend).  The loaded compare value must be compared against a
truncated smaller type, so down-cast appropriately for each size.

The issue is apparent on 64-bit machines with code, such as
atomic_dec_unless_positive(), that sign-extends from int.

64-bit machines generally don't use the generic cmpxchg but
development/early ports might make use of it, so make it correct.

Signed-off-by: Matt Evans <mev@rivosinc.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>