Support dedicated square root operator

[huningxin]

Use case and sample models

This op was proposed by @fdwr (Thanks Dwyane!) as a WebNN v2 op in WebML WG Teleconference – 29 June 2023 to support transformer models, such as segment anything model.

Cross-framework support

It is widely supported by ML frameworks, such as

• TensorFlow supports tf.math.sqrt.
• Pytorch supports torch.sqrt.
• ONNX supports Sqrt.

Cross-platform implementability

The dedicated square root operator is widely available in native ML APIs / libs, e.g.,

• XNNPACK xnn_define_square_root
• DirectML DML_ELEMENT_WISE_SQRT_OPERATOR_DESC
• MPSGraph squareRootWithTensor.

Other references

Currently the Chromium CL-4615724 maps pow(x, 0.5) to xnn_define_square_root for WebNN XNNPACK backend. It could be exposed via the dedicated square root operator when WebNN spec supports it.

/cc @wacky6

[fdwr]

Motivations include:

• Backends often have a more optimized version of it available than via pow.
• It avoids needing to create and upload tiny temporary tensors (of 0.5).

Note any backends lacking a dedicated square root can still trivially achieve it via the fallback pow(x, 0.5).

[inexorabletash]

This can be closed off c/o PR #478, correct?