Low precision accumulation flag

[nikalra]

• Adds support for setting MLModelConfiguration.allowLowPrecisionAccumulationOnGPU

[TobyRoseman]

This is a great feature! Thanks for the pull request.

The code itself look good, but there are a few things that I think should be addressed:

1 - I think we need to be explicit (verbose even) in the naming of this flag in our public APIs. Let call it allow_low_precision_accumulation_on_GPU.

2 - I’d like to see new unit tests created for this functionality. Rather than tacking it on to existing unit tests. Ideally these tests would load the model normally and load model using this flag. Then verify that the results are close.

3 - I think if a user sets this option to True but is also using a compute unit that doesn’t allow GPU, then they should get a warning.

I have kicked off a CI run to this change.