[codegen] Fold uextend.i64.i32 into a no-op whenever possible on x86

[bnjbvr]

Still a WIP. This removes spurious moves (generated by uextend.i64.i32) whenever the previous instruction did zero the upper 32 bits for us.

Remaining things to do:

• Fix constraints of the null recipe being not satisfied.
• Make a whitelist of (instruction, encodings) that are acceptable, so it handles most of the cases where this optimization is useful.
• Once Improve extractlane and replacelane on x86 #943 is merged, add back the uextend in the wasm generation of ExtractLaneU, and fold it back in this new peephole optimization pass.

[lars-t-hansen]

This is a fine stopgap solution. In general however we want something table-driven (ie a machine-generated automaton) that runs during machine code generation and acts on the actual encodings emitted. Such an automaton can likely subsume other optimizations, eg the shrinking pass.

[bnjbvr]

I'll need to re-work this a bit, and think about it (maybe publicly as part of an issue). Closing since there are no good reasons to keep this WIP PR open for too long.