Move some fast invoke checks to CanUseMterp

This speeds up arm64 golem interpreter benchmarks by 1.5%.

Test: test.py -b -r --interpreter --host
Change-Id: Ia9d7c885cd488de56c6b726373072070b509bdf1