arm64: improved arch_switch() implementation

Make it optimal without the need for an SVC/exception  roundtrip on
every context switch. Performance numbers from tests/benchmarks/sched:

Before:
unpend   85 ready   58 switch  258 pend  231 tot  632 (avg  699)

After:
unpend   85 ready   59 switch  115 pend  138 tot  397 (avg  478)

Signed-off-by: Nicolas Pitre <npitre@baylibre.com>