arm64/nterp: Free up registers xLOADX and xLOADD...

... for other uses. Adjust xLOADW/xLOADS by a fixed offset
to calculate the wide register handler addresses.

Local benchmarking of nterp-to-non-nterp calls on Pixel 3
shows minor changes for big cores (from 2% improvements to
2.5% regressions for `float` benchmarks, from 1.5%
imorovements to 1% regression for `double` benchmarks) and
litte cores (0%-4.5% improvements for `float`, 0%-1%
regression for `double`).

Test: testrunner.py --target --64 --interpreter
Flag: EXEMPT PURE_REFACTOR
Change-Id: I918dfad121694fab1db0e6fe42f0d7e6d4e5caf4