"Unintern" DataInstForm (inlining its fields back into DataInstDef).

[eddyb]

Effectively undoes:

• Intern, as DataInstForm, the kind and output_type fields of DataInstDef. EmbarkStudios/spirt#28

Some quick unscientific testing reveals no significant perf impact (i.e. the difference is lost in the noise).

The motivation for undoing this interning is the prospect of combining DataInstDef into NodeDef (as mentioned in #7), and for unrelated pragmatic reasons, NodeDef can't have its outputs interned (though long-term maybe we could intern the kind field, if really necessary, assuming we first take "child regions" out of it).

The one thing I realized too late there's no pre-existing consensus for, is the output_type, which used to be between kind and inputs (matching DataInstFormDef, in fact), while NodeDef's outputs is the last field.

The only argument to have outputs first is the let (out0, out1) = foo(in0, in1); style syntax (though pretty-printed SPIR-T omits the let), but I'm not sure that's worth flipping the order over.

[LegNeato]

Don't understand the code yet, but the changes look pretty mechanical and clean stuff up nicely.