Use "dynamo" to rewrite int(...), float(...), and bool(...)

[makslevental]

This PR uses https://github.com/makslevental/pyframe_eval to implement our own version of dynamo and support int(...), float(...), and bool(...), i.e., a custom stackframe evaluator that "just-in-time" rewrites the ast such that

class MyIntModule(nn.Module):
    def forward(self, x):
        y = int(x)
        return y

becomes (roughly)

class MyIntModule(nn.Module):
    def forward(self, x):
        y = pi.pi_int(x)
        return y

which then gets lowered to

func.func @forward(%arg0: !torch.vtensor<[1],si64>) -> !torch.int {
  %0 = torch.aten.Int.Tensor %arg0 : !torch.vtensor<[1],si64> -> !torch.int
  return %0 : !torch.int
}

Why go through all of this trouble when a decorator would accomplish the same thing? Because this will handle all follow-on calls from within forward too (e.g., some helper that we can't decorate).

Just to be very clear: there's still no PyTorch in the mix here, I simply implemented the same thing that they did (and packaged it seperately in pyframe_eval) and instead of custom interpreting the bytecode of the stackframe, I made it rewrite the ast of the stackframe.

This and (pyframe_eval) is still not fully tested (so it might segfault or assert inside of cpython) but it basically works:

total # of tests: 930
total # of skips: 18
total # of passes: 741
total # of xfails: 16
total # of exceptions: 90
total # of lower to linalg failures: 25
total # of ir differences: 40

cc @powderluv

[makslevental]

this still has to be figured out

[makslevental]

this is the key

[powderluv]

Can we copy a version of it into PI ? Also you probably want to set the license for that. Ideally Apache 2.0 is best for PI.

[makslevental]

you/we/I can copy but do you really want it here? It's hairy cpython internals - might as well pawn maintenance off on me into perpetuity.

[makslevental]

okay I licensed it WTFPL (not specifically for you but you know just for everyone...)

[makslevental]

@123epsilon this isn't done yet...