keepDimensions of reduction ops

[lisa0314]

XNNPACK can't support keepDimensions = true for ReduceMean.
This issue was raised by @huningxin in WebNN Chromium CL review. Thanks Ningxin!

[fdwr]

keepDimensions just controls the graph builder's output dimension description and does not change how tensor elements are read or written (the output memory is identical either way). So a backend supporting keepDimensions or not itself doesn't matter for the WebNN caller, since it's just a matter of updating the tensor description in the backend - a nop reshape.

For contrast, the DirectML backend only supports the reverse, keepDimensions == true, but keepDimensions == false still works fine because the lightweight tensor dimension is massaged before calling DirectML. In both cases below, DML_OPERATOR_REDUCE receives [2,1,4] for the output tensor desc, because the backend inserts the dummy 1's (note there is no actual "reshape" here, but conceptually it is).

1. input sizes = [2,3,4], keep dimensions = true, axis = 1, output sizes = [2,1,4]. DML sees output [2,1,4]
2. input sizes = [2,3,4], keep dimensions = false, axis = 1, output sizes = [2, 4]. DML sees output [2,1,4]

Conversely, XNNPack should see:

1. input sizes = [2,3,4], keep dimensions = true, axis = 1, output sizes = [2,1,4]. XNNPack sees output [2,4]
2. input sizes = [2,3,4], keep dimensions = false, axis = 1, output sizes = [2, 4]. XNNPack sees output [2,4]

[inexorabletash]

Based on @fdwr's comment, would this be resolved by adding a non-normative note with hints to implementers?

[huningxin]

XNNPACK now supports keepDimensions = true for ReduceMean: google/XNNPACK@5ecf0769c54cd224bd0026fe2c