Dart bindings for ExecuTorch, plus a parity gate over the exported goldens

[tonytonycoder11]

I built a Dart and Flutter layer over ExecuTorch, out at 1.0 last week:
https://github.com/NaCode-Studios/Fluttorch (Apache 2.0)

The bindings are the boring half. The other half is that fluttorch-export wraps torch.export
and writes the .pte, a manifest, and reference outputs from real validation inputs, all
together. Those outputs replay in the Dart test suite, and the tolerance comes from the recipe
and precision the manifest recorded instead of a number picked by hand:

FAIL  parity/case-3
      backend: xnnpack  quantization: int8-static
      output "logits"  max |Δ| 1.72 > Tolerance(atol 0.2, rtol 0.2, cos ≥ 0.998)
      no layer attribution: backend "xnnpack" offers no activation taps

That gate is why I wrote my own dart:ffi binding rather than using an existing one. It needs a
backend pinned at load, repeatable execution, activation taps, and output buffers the caller owns.

Two questions:

1. Is there a supported way to read intermediate activations under XNNPACK? Right now I can only
   report which output moved, not which layer.
2. The parity matrix treats Core ML and MPS as float16 by default and gives each its own tolerance
   column. Is that the right default, or is there a lowering flag I've missed?

Separately: I have a model that lowers fine and then fails to execute, identically under
ExecuTorch's own Python runtime. Worth opening an issue with a repro?