tblgen: add ResultIndex primitive

[Pangoraw]

to refer to the result value of an operation in the tablegen rules.

cc @mofeing, as we had discussed about this.

[Pangoraw]

ideally mincut can also identify equivalence to merge nodes in the graph but this can still be interesting to write more memory optimized rules

[wsmoses]

I'm slightly worried about this [though less so for EnzymeMLIR than Enzyme proper] because the analysis to detect if a value is needed for the reverse looks at users of the value, and doesn't consider "itself" as a potential user. In principle is there a reason why re-running the op wouldn't suffice (and CSE-style optimizations won't make it equivalent perf?]

[Pangoraw]

CSE-style optimizations won't make it equivalent perf

the push / pop mincut does not currently treat the recomputation as equivalent and as such will always choose to cache the operand of the original op instead of its result.

scf.for {
  enzyme.push %v
  %0 = math.exp %v
  ...
}

scf.for {
  %v = enzyme.pop
  %dx = math.exp %v
  ...
}

could be

scf.for {
  %0 = math.exp %v
  enzyme.push %0
  ...
}

scf.for {
  %dx = enzyme.pop
  ...
}

because the graph looks something like this

               %v
              /  \
math.exp (fwd)    math.exp (bwd)
             |    |

so it is less costly to store %v than its result

[wsmoses]

That makes sense, though I feel like we should min cut here (and that would regardless be useful for other things too)