Convert ci-shark-ai.yml to pkgci_shark_ai.yml.

[ScottTodd]

This builds on #589 to make progress on #584.

On the surface, this adds complexity right now. The benefits are:

• Integration test workflows use dev packages, without needing to build them from source or use editable installs
• Users and developers can access the same dev packages to reproduce CI results
• We only need to build shortfin from source once for multiple jobs (potentially saving 2-3 minutes)
• Only one runner needs the build requirements (potentially including clang, ninja, CMake, Rust, etc.), other runners only need Python

This also switches to using Python venvs, tightening configuration control in workflows and ensuring that only the expected packages are installed and used for tests, without leaking installed packages across workflow runs. That does increase the workflow time though.

---

Logs before: https://github.com/nod-ai/shark-ai/actions/runs/12058325322

• Total: 7m6s
• 3m5s on Install pip deps (no venv, reusing already installed packages and getting the wrong IREE versions)
• 3m14s on Run LLM Integration Tests

Logs after (cold cache): https://github.com/nod-ai/shark-ai/actions/runs/12059301876/attempts/1?pr=625

• Total: 9m57s
• 2m42s on Build shortfin (cold cache)
• 2m37s on Setup venv (from scratch, no deps reused)
• 16s on Install nightly IREE packages
• 3m25s on Run LLM Integration Tests

Logs after (warm cache): https://github.com/nod-ai/shark-ai/actions/runs/12059301876/attempts/2?pr=625

• Total: 8m20s
• 1m24s on Build shortfin (warm cache, 46% cache hit rate)
• 2m23s on Setup venv (from scratch, no deps reused)
• 17s on Install nightly IREE packages
• 3m23s on Run LLM Integration Tests