The baseline implementation of WebNN ops

[huningxin]

To establish the baseline of WebNN conformance test and answer the requirements raised in WebML WG Teleconference – 2 Dec 2021, including

by @wchao1115

all computation is done in double precision
we don't want any intermediate casting
an open source ref, anyone can look at the code and be confident

by @dontcallmedom

the codebase should be easy to review, not too many layers of abstraction

I would propose to create a baseline implementation of WebNN ops that:

• Uses JavaScript double precision calculation
• Uses plain algorithm without depending on third-party library
• Hosts on group repo, such as https://github.com/webmachinelearning/webnn-baseline

I have been working with @BruceDai on a staging repo: https://github.com/huningxin/webnn-baseline, with some initial ops, like conv2d. If it looks like a right approach, we are happy to contribute it to the group repo for wider review.

/cc @anssiko @wchao1115 @dontcallmedom

[huningxin]

As this is for w-p-t, mark it with 'cr'.

[anssiko]

Initial implementation: webmachinelearning/webnn-baseline#1