More complete opaque type support

[junrushao-openai]

Noticed there are a few features missing for opaque support:

• Structural equality with opaque feat(core): Support eq_s and hash_s for opaque objects #74
• Prioritize opaque types once registered over other types (e.g. callable) feat(core): Support eq_s and hash_s for opaque objects #74
• Serialization & deserialization with opaque feat(core): Support serialization of opaque objects #75
• Deep copy with opaque feat(core): Support deepcopy in opaque #76

[potatomashed]

What are the usecases of embedding arbitrary python classes into a compiler IR? MLC does support that as workarounds, but still, it's usually against a compiler's design philosophy to have any opaque nodes inside an IR.

I don't believe other compiler infra, e.g., MLIR or any other infra would support that easily. If it's really used, it's possible that the compiler can be designed better.

A few alternatives:

• Have a special "opaque" node in the IR and have them as inputs of functions - this way, you don't have to embed external classes in the IR;
• Use intrinsics to to represent opaque - this way, opaqueness is embedded in intrinsic calls instead of directly inside an IR

[potatomashed]

#74 brings some support; serialization will be supported in follow up PRs.

[potatomashed]

#75 gives serializability. By default via jsonpickle but can be overridden.

[potatomashed]

#76 gives deepcopy