[onnx] Enable f64 backends for onnx test suite

[rsuderman]

For rocm cpu and cuda we can support f64 types for the onnx backend. It makes more sense to support testing using these types on backends where it is available.

[ScottTodd]

false is the default right? You can just leave the flag off then.

[rsuderman]

It is not. I have to set to false to disable.

[ScottTodd]

iree/compiler/src/iree/compiler/Pipelines/Options.h

Lines 53 to 58 in 6c45bef

	// Gate various type based demotion passes that run before anything else.
	bool demoteI64ToI32 = false;
	bool demoteF32ToF16 = false;
	bool demoteF64ToF32 = true;
	bool promoteF16ToF32 = false;
	bool promoteBF16ToF32 = false;

ah. Weird that we have it explicitly set to false true in so many places then.

[ScottTodd]

Test results appear mixed on https://github.com/iree-org/iree/actions/runs/10170532539/job/28130212435?pr=18056. Some newly passing, some newly failing to compile, some newly crashing while compiling :P.

I think we should just test with default flags, whatever those may be. The flags that change types at the input stage are generally sketchy though, so it isn't too surprising if enabling/disabling them changes behavior in unexpected ways.

[rsuderman]

Once llvm/torch-mlir#3579 merges in and sync (along with a llvm bump). We should see an increase in testing coverage.

Its worth noting some of our failures are actually passing, the rtol is just too small for some tests (e.g. onnx.Pow)

[ScottTodd]

Its worth noting some of our failures are actually passing, the rtol is just too small for some tests (e.g. onnx.Pow)

We had a discussion a while ago about updating the test suite(s) to support changing tolerances for individual tests / backends. Never followed up on that, I don't remember who was looking at it.