local_irq_enable() may expand into very different code, so it rather should
stay in C.  Also this keeps the assembler code size constant which keeps
the rollback code simple.  So it's best to split r4k_wait into two parts,
one C and one assembler.

Finally add the local_irq_enable() to r4k_wait to ensure the WAIT
instruction in __r4k_wait() will work properly.
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>