Use variable-length encodings for REX encodings

[bnjbvr]

Variable-length encodings (introduced to support the proper machine code for loads/stores which memory operand is in a given register on x86) could allow removing the need to have both the REX vs non-REX encodings.

This would have a few benefits:

• reduce the number of encodings in the legalizer interpreter, making it faster
• allow some encodings to accept REX registers, when they couldn't before (at least regmove is affected by this issue)
• reduce the number of methods in PerCpuModeEncodings

[bjorn3]

It would also make the instruction shrinking pass unnecessary when you want to avoid unnecessary REX prefixes. That would for example avoid https://github.com/bjorn3/rustc_codegen_cranelift/issues/146#issuecomment-449474527 when using opt_level=none.

[bjorn3]

The new backend framework emits a REX prefix only when it is necessary.