Add `--verify-roundtrip` on `substrait-opt` tests

[mortbopet]

We should consider adding a tool substitution for substrait-opt to always do roundtrip checking for any test that we do - see e.g. in CIRCT.
I've personally found this to be quite useful, catching errors which weren't on my radar when "just" trying to test some pass using an opt tool.

[ingomueller-net]

Yeah, good idea! What I don't know yet and haven't done in #109 is whether we can do a tool substitution that is a chain of commands. (We can, of course, build something into substrait-opt or substrait-translate, like in CIRCT.)

Also, currently, google.protobuf.Any messages don't deserialize well to JSON due to a (presumed) bug in the Protobuf library, see protocolbuffers/protobuf#20675. So plans that use Substrait extension currently don't round-trip just because of that.

Thinking out loud: maybe we don't need to roundtrip through JSON? Maybe we can implement a --verify-roundtrip by comparing (1) the parsed Protobuf message converted to textformat with (2) the same message imported into MLIR, exported back to Protobuf textformat.

I wonder whether that isn't quite a bit more elegant than #109...

[ingomueller-net]

You may have noticed but --verify-roundtrip doesn't test the same type of roundtripping as #109 does. The former tests wheter MLIR IR -> custom assembly -> MLIR IR is the identity. The latter tests JSON -> Protobuf -> MLIR IR -> Protobuf -> JSON is the identity.

That doesn't change much for the general idea, though. We can add our own --substrait-verify-roundtrip to substrait-translate and test round-tripping based on the Protobuf formats.

[ingomueller-net]

Independently, we should add the substrait-op --verify-roundtrip tests for our dialect tests.

[mortbopet]

It's definitely two separate levels of roundtripping; whether or not we could (or should) build in some automated mechanism to verify everything though not just IR roundtripping but also import/export is a good question.