Add oneDNN/XNNPACK backend and support MobileNet/SqueezeNet model

[fujunwei]

The commit supports 5 first-wave ops required by MobileNet/SequeezeNet example including Clamp, batchNormalization, LeakyRelu, Concat and Gemm.

It also adds XNNPACK and oneDNN backends that is mainly for the investigation of WebNN Operation Level Execution
use case. So only a limited set of tests (such as of conv2d) is expected to pass.

It also adds experimental sync build and compute API to implement TF.js Wasm backend by WebNN API.

[huningxin]

Great work. Thanks much @fujunwei . The implementation would help the group for spec discussions, in particular for recent wasm op-level execution webmachinelearning/webnn#156.

I left some comments for clarity. Please take a look.

[fujunwei]

Thanks, done. Please take a look again.

[BruceDai]

Thanks much @fujunwei!

I tested with this PR by both DML backend and OpenVINO backend, here's tests and examples PASS status table as following:

Backend	End2EndTests (Pass / Total)	LeNet Example	SqueezeNet1.1 nchw Example	SqueezeNet1.0 nhwc Example	MobileNetV2 nchw Example	MobileNetV2 nhwc Example
DML	148 / 148	✔	✔	✔	✔	✔
OpenVINO	146 / 148	✔	✔	✔	✔	✔

And two MaxPool2d tests with dilations options (Pool2dTests.MaxPool2dDilationsDefault, Pool2dTests.MaxPool2dDilationsNhwc) failed when running OpenVINO backend. @fujunwei PTAL, thanks.

[huningxin]

LGTM with a nit. Thanks @fujunwei for contributing this PR and @BruceDai for testing.

[fujunwei]

Thanks @BruceDai, current OpenVINO doesn't support dilated pooling, and only MaxPooling support dilation in DML. I filed a issue-180 to discuss if it should be defined in WebNN Spec.

[BruceDai]

Thanks much @fujunwei for kindly explanation!

[huningxin]

It also adds XNNPACK and oneDNN backends that is mainly for the investigation of WebNN Operation Level Execution use case. So only a limited set of tests (such as of conv2d) is expected to pass.

It also adds experimental sync build and compute API to implement TF.js Wasm backend by WebNN API.

Given the PR of sync API was merged and the investigation of wasm op-level execution was done, I am proposing to merge this PR to allow participants to reproduce the results. And it would also unblock the development for #6 and #12 .

Any objections? @anssiko @wchao1115 @pyu10055

[anssiko]

SGTM!