Unbundling oneDNN

Breaking out the discussion from #175:

> @njzjz: Currently tensorflow here is not built against MKL (OneDNN), but do we plan to?

> @xhochy: Tensorflow is building oneDNN from source in the build currently. That is probably the first 1-2h of CI. Getting that from the feedstock instead of building from source would help a lot.

I looked at how upstream builds this, and - annoyingly - it's a bit of a mess. There seem to be three different versions of onednn being [pulled](https://github.com/tensorflow/tensorflow/blob/v2.7.0/tensorflow/workspace2.bzl#L177-L208) in, with one of them (0.21.3) being quite [old](https://github.com/oneapi-src/oneDNN/releases/tag/v0.21.3) and not [available](https://anaconda.org/conda-forge/onednn/files?sort=time&sort_order=asc) in conda-forge.

I'm happy to start a PR that tries to rip things and see how far we get, but just wanted to get an opinion (@xhochy?) if replacing the upstream version is even realistic given the different versions? Or maybe we just remove the most important one?

Start by reading the upstream TensorFlow references in tensorflow/workspace2.bzl, especially lines 177-208, and compare the three oneDNN versions with the conda-forge package availability linked in the issue. Done means establishing whether the feedstock package can replace the source-built dependencies without incompatible version requirements or identifying which upstream dependency can be removed.

Written by the indexing model from the issue text.