Bump cmake version for GPU build

[t-vi]

The cmake version (3.10) in Ubuntu 18.04 does not cope well with the
more advanced cmake use in libtorch surrounding the CUDA target.
We switch to a self-built cmake 3.14 (already used by arm and i386 CI).

The context for this is #10758 .

Thank you @masahi @driazati for your helpful discussion, errors are my own.

[masahi]

@Mousius Seems TF-aarch64 v2.6.2 we try to install at https://github.com/apache/tvm/blob/main/docker/install/ubuntu_install_tensorflow_aarch64.sh#L29 is no longer available? See https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-11156/1/pipeline/

Maybe we should switch to TF 2.7?

[Mousius]

Looks like https://snapshots.linaro.org/ldcg/python-cache/tensorflow-aarch64/ is down for some reason, I've reached out to find out why.

The reason for 2.6.2 is to maintain parity with ci_cpu:

tvm/docker/install/ubuntu_install_tensorflow.sh

Lines 23 to 26 in 9fd279b

	pip3 install \
	"h5py==3.1.0" \
	keras==2.6 \
	tensorflow==2.6.2

Upgrading to 2.7.0 should really be done in both places?

CC @gromero @leandron

[masahi]

@driazati For a PR like this where only docker/Dockerfile.ci_gpu (or other install scripts used by subset of nodes) is modified, can we skip building ci image for unrelated nodes? Docker scripts update has been painful because "Docker Image Build" stage in the pipeline can easily fail for strange reasons and especially frustrating when the failed image is for unrelated nodes...

[Mousius]

@t-vi / @masahi it looks like https://snapshots.linaro.org/ldcg/python-cache/tensorflow-aarch64/ is back, I've restarted your build in Jenkins to see if we can get this moving 😸

[t-vi]

Thank you, @Mousius !