Revisit interpret loop optimizations

[edubart]

Context

The interpreter loop got about ~10% performance regression between emulator 0.19 and 0.20. This happened after the refactoring we made in the interpreter to make the TLB code more sane and encapsulated access to address offsets through host_addr type.

Possible solutions

Investigate what assembly instructions were added by tracing the host amd64/arm64 instructions when running the interpreter between emulator releases. Then come up with a code refactor to save host instructions in the interpreter hot loop while retaining code clarity and sanity (no crazy hacks for performance).

[edubart]

This was solved, the culprit was the removal of the early XOR check using fetch cache when during instruction fetch, which was re-added.