Container images: Support "template" images and use it for cuda-tests

[copybara-service]

Container images: Support "template" images and use it for cuda-tests

The gpu/cuda-tests and gpu/cuda-tests-12-8 are duplicates of each other,
other than the CUDA version they use. As more CUDA versions are added, this
isn't a scalable pattern.

This change adds support in the Makefile (more specifically
tools/images.mk) for "image templates" and "image template instances".
If an image directory ends in .tmpl, it will be ignored from the set of images
that the Makefile recognizes. Instead, this directory can be used to
instantiate other images.

For example, given the following filesystem structure:

images/
├─ my-little-image.tmpl/
│  └─ Dockerfile
├─ my-little-image.foo.bar → my-little-image.tmpl (symlink)
└─ my-little-image.baz.qux → my-little-image.tmpl (symlink)

Then this will effectively create two images, my-little-image.foo.bar and
my-little-image.baz.qux. It will not create a my-little-image.tmpl image.

The behavior of the template instance images is determined by the
TEMPLATE_VERSION build argument passed to my-little-image.tmpl/Dockerfile.
This argument takes on the value of everything after the first . character of
the last component of the template instance image name. For example, the image
my-little-image.foo.bar will be built with docker build --build-arg=TEMPLATE_VERSION=foo.bar, whereas the my-little-image.baz.quux
will be built with docker build --build-arg=TEMPLATE_VERSION=baz.qux. The
my-little-image.tmpl/Dockerfile image definition file can use this variable to
make the necessary tweaks to distinguish these two images.

#codehealth