Consistently support compound equality in the IR

While working on #380 it has become clear that supporting compound equality (arrays, tuples and combinations of them) is not just a JIT problem.

For example, optimizations like arithmetic simplification assume the arguments of Eq comparisons in the IR are just simple bit vectors: https://github.com/google/xls/blob/24046e424af962efd2b151d383cb9e01071bc1c8/xls/passes/arith_simplification_pass.cc#L518

Similarly, the IR verifier currently doesn't allow Eq comparisons on non-bit things: https://github.com/google/xls/blob/24046e424af962efd2b151d383cb9e01071bc1c8/xls/ir/verifier.cc#L507

The reason we can get to the JIT at all with a compound equality is that test constructs aren't consistently converted to IR and verified.

We should have a consistent story about how to handle compound equality in the IR. Options include

1. Always desugaring compound equality as part of converting to IR. (Pro: Simplifies IR handling downstream, Con: Desugared equality means we lose high-level structure and could turn into more complex and less readable generated Verilog).
2. Requiring all IR tools to hand compound equality (Pro: We don't lose the underlying equality structure. Con: Extra complexity in everything that handles equalities in IR).
3. Context-dependent switching between desugaring or not (maybe via a flag, maybe based on what downstream consumers are). Pro: Can do the most convenient thing for different use cases. Con: Extra complexity (especially because some IR invariants become weaker).