Originally, this task would download all artifacts requested in the
input file, check them all, and then upload them all to the registry in
one invocation of "oras push".

This had two problems. First, if "oras push" flaked out part way
through and the user needed to retry their pipeline, the entire download
section would need to be run again needlessly. Second, for extremely
large artifacts with lots of medium-sized files, an enormous PVC would
be needed to hold all of them between download and push to the registry.

The change here addresses both problems.

First, files are downloaded, checked, pushed to the registry and then
deleted from local storage - one at a time. This obviates the need for a
large volume to store all files at once, since only enough storage is
needed to store one file, not all of them.

Second, as files are considered, first the registry is checked to see if
the blob has already been pushed there. If it has, then skip the
download step. This has the effect of greatly improving the runtime for
artifacts where only one or two of many files have changed since the
last taskrun.

Theoretically, this works if the IMAGE reference contains a port number.

Co-authored-by: Adam Cmiel <[email protected]>