Split pkgci_shark_ai.yml for (short / long-running) X (cpu / gpu)

[renxida]

Since we have pkgci and build packages once before running tests (#780), we can now build the packages once and download them for running multiple tests in parallel.

In particular, it would be nice to have GPU tests for shortfin, as well as a split between quick tests that run small / toy models, and slow test that run full-sized models.

This also enables parallel execution between e.g. open Llama and meta llama.