[StackIR] Optimize the simple case of a multivalue tee and extracts

[kripken]

When getting a multivalue from e.g. a call and then consuming it immediately,
we get this pattern:

      (tuple.extract 3 0
        (local.tee $temp
          (call $multivalue-return)
        )
      )
      (tuple.extract 3 1
        (local.get $temp)
      )
      (tuple.extract 3 2
        (local.get $temp)
      )

We save the entire tuple to a local, then read index by index. In StackIR
we can just remove all the tuple operations, as the value is on the stack
and ready to be used.

Fixes #8618

[tlively]

Might be interesting to add round trip output, too, to make sure the binary writer still does the right thing.

[kripken]

Already done 😉 see the command above

[tlively]

Oh, I see. It's too bad we can't get a text output that matches the written binary exactly (i.e. with the tuple pseudoinstructions lowered away).

[kripken]

We can get close without the roundtrip - the StackIR is printed right before binary emitting. I pushed a change now to include both outputs (with a different lit prefix).