Stop passing CMAKE_TOOLCHAIN_FILE, append to CMAKE_FIND_ROOT_PATH

[mgorny]

Checklist

• Used a personal fork of the feedstock to propose changes
• Bumped the build number (if the version is unchanged)
• Reset the build number to 0 (if the version changed)
• Re-rendered with the latest conda-smithy (Use the phrase @conda-forge-admin, please rerender in a comment in this PR for automated rerendering)
• Ensured the license file is being packaged.

---

Stop passing obsolete CMAKE_TOOLCHAIN_FILE for CPU builds.  Instead,
append the source directory to CMAKE_FIND_ROOT_PATH directly.  Do that
unconditionally, to improve consistency between CPU and GPU builds,
and hopefully make it possible to build MKLDNN with CUDA enabled.

---

I haven't bumped the build number since this shouldn't be changing anything (and at least at a first glance, I didn't notice any differences). It's getting late today, so I'm going to look at whether this actually fixes building against vendored mkldnn tomorrow (I've originally tested only against my system-mkldnn branch).

[mgorny]

And thanks to @isuruf for explaining the toolchain thing to me.

[conda-forge-admin]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe/meta.yaml) and found it was in an excellent condition.

[hmaarrfk]

I haven't bumped the build number

If this is the case, I would rather you set the time limit to 1 min to effectively force the CI to not publish at all. I am a little concerned about the GPU build since it doesn't show up on the CIs.

[mgorny]

I haven't bumped the build number

If this is the case, I would rather you set the time limit to 1 min to effectively force the CI to not publish at all. I am a little concerned about the GPU build since it doesn't show up on the CIs.

I'm sorry, I didn't know I can do that. In fact, I haven't found any information on that yet. Perhaps I could use [ci skip] ***NO_CI*** in the future instead?

Also, I'm currently confirming that this fixes USE_MKLDNN=1 with CUDA, per #233. Would you like me to add enabling it to this pull request, or wait for it till we unvendor it (per #289) instead?