MIPS: Loongson64: smp: Correct nudge_writes usage

Previously every write to SMP regisers are followed by nudge_writes,
this incures a huge performance penalty because nudge_writes involves
SYNC, which will be globalized on chip.

Only set off nudge_writes when we really want other cores to see the
result ASAP. Also replace read/write functions to relaxed version because
we don't need extra barriers to protect against DMA.

Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
Signed-off-by: Thomas Bogendoerfer <tsbogend@alpha.franken.de>