change upstream to ROCmSoftwarePlatform/tensorflow-upstream

[wuxxin]

• use branch+commit from docker.io/rocm/tensorflow@lastest at 2022-08-03
• updates tensorflow to 2.9.2

rationale for switching upstream:

• the docker image of rocm/tensorflow uses ROCmSoftwarePlatform/tensorflow-upstream

  • https://hub.docker.com/layers/tensorflow/rocm/tensorflow/latest/images/sha256-2f92080ae9c923c89b1a25cbecec95c2de312d111bbe442d5aa1388c6c2aec44?context=explore

• the rocm build readme of google/jax uses ROCmSoftwarePlatform/tensorflow-upstream

  • https://github.com/google/jax/blob/main/build/rocm/README.md

Warning:
Although ROCmSoftwarePlatform/tensorflow-upstream is a maintained fork, it has currently around 3600 additional commits for ROCM support, so the code is vastly different, and probably prone to different errors than original upstream

[acxz]

For this tensorflow-rocm package, we want to use the upstream (tensorflow/tensorflow) source as eventually this PKGBUILD will be merged with the archlinux tensorflow package.

This does cause us trouble in that tensorflow/tensorflow tends to be farther behind than ROCmSoftwarePlatform/tensorflow-upstream.

I'm going to have to reject this PR on the above rationale. However, it seems like tensorflow/tensorflow is pretty much always further behind, rationalizing using the amd fork of it. I think the way to move forward with this, would be to create a separate PKGBUILD that uses the amd source, maybe calling it tensorflow-amd-git? This should appease users that just want to get up and running with tensorflow on rocm. I would also need to make sure that tensorflow with this new source builds in a clean chroot, for fear of creating more confusion and unnecessary bug reports by adding another package that doesn't build. Any thoughts?

Just to clarify the purpose of this PKGBUILD is to build upstream tensorflow with rocm and test it for long enough so that it can finally be added into the PKGBUILD for the official Archlinux package. By changing the source, we cannot accomplish this purpose.

[wuxxin]

make sense for the goal to merge into archlinux tensorflow to not have different upstreams.
the AMD fork and upstream will probably converge at some point in time.

in the meantime i renamed and registered tensorflow-amd as alternative.