Fix dimensions for 0D scalars

[fdwr]

From issue: #390

A 0D tensor shape vs a 1D tensor with a single element are semantically and functionally distinct, but the current spec mistreats 0D scalars and 1D tensors like they are the same thing, and this messes with the shape inference for operators like reduceSum (with all axes and keepDimensions = false) or gather where the output rank depends on the correct rank for input and indices tensors. This caused models like Segment Anything and Stable Diffusion to fail while prototyping, and the issue arose again during Chromium and ORT WebNN EP code review.

[1,2,3] - 3D tensor shape
[1,2] - 2D tensor shape
[1] - 1D tensor shape
[] - 0D tensor shape (scalar) <--- (every known ML library represents 0D scalars via the shape [])

Fixing this in the Chromium fork needed changing just two lines, and we want to address this sooner than later before changing just a few lines becomes multiple lines.

---

Preview | Diff

[zolkis]

Thanks for spotting / correcting that.

[anssiko]

@fdwr noticed while checking the PR queue this needs a rebase. Good discussion in this PR should be factored in too I believe.

[inexorabletash]

Other things that should be rolled into this PR:

• In constant() "Set descriptor.dimensions to undefined." - should set it to an empty list instead. And the note "In the case of a scalar constant, descriptor.dimensions is ignored." should either be removed or updated; it's not that dimensions is ignored, it's that a rank of 0 defines it as being a scalar.
• validate MLOperand has "If desc.dimensions exists" - remove that clause?
• lstm() accesses dimensions[0] without a size / rank check first (in step 2) - add a check? I couldn't spot any other instances.
• Make the default for MLOperandDescriptor's dimensions be []

[fdwr]

There's a build error (No 'idl' refs found for 'MLCommandEncoder') presumably from https://github.com/webmachinelearning/webnn/pull/546/files. I'll leave the status text there but just unbracket it to hopefully appease bikeshed.

Build completes now. If I should remove status text containing MLCommandEncoder entirely instead, let me know.

[inexorabletash]

LGTM!

[zolkis]

Thanks, looks good!