Cache whether we should use interpreter invoke fast-path.

There is many conditions to check on each invoke.  If they all pass,
cache the result in the access flags of the called ArtMethod.

This speeds up arm64 golem interpreter benchmarks by 4%.

Test: ./art/test.py -b -r --interpreter --host --64
Change-Id: I86313dbdba84cbf8c707aae84e0ff05b6e034700