CICD: Runs a full GPU install on an EC2 instance

[robotrapta]

Pretty big (but useful!) change. Adds a GHA workflow step that in a fully automatic way:

• Creates a new g4 EC2 instance
• Installs EE on it. (Installs K3s, and installs our YAMLs.)
• Checks that the EE install script runs successfully
• Checks that the k8 deployments come up
• Checks that the SDK can do minimal things through the EE (whoami, list-detectors)
• (Does not tears down the EC2 infra - there's a sweeper which will run every 30 minutes and do that async so the pipeline doesn't have to wait for it, because it takes ~7 minutes to turn off a G4 instance.)

It uses pulumi for infra, and relies on pulumi infra defined in the GL_Public AWS account.

It does NOT yet:

• Check that any inference models work.

(Note: there is replication of #169 which improves the workflow YAML and validation thereof directly.)
(Note: this relies on resources defined in the gl_public account defined in the internal ci-infra repo.)

[robotrapta]

Okay I think this is actually working again and really ready for review! I'm gonna deliberately break the k8 YAML and make sure the test fails (it wouldn't have before) and then restore it.

[robotrapta]

It's spec not shrek :)

Error from server (BadRequest): error when creating "deploy/k3s/edge_deployment/edge_deployment.yaml.tmp": Deployment in version "v1" cannot be handled as a Deployment: strict decoding error: unknown field "shrek"

[tyler-romero]

Oooohhh

[tyler-romero]

Is this a worry b/c dependabot wont check uv.lock files yet?

[robotrapta]

It's not ideal. But this dependency closure is pretty tiny - only a handful of packages in here. And it looks like dependabot is going to add uv this quarter.

[tyler-romero]

I mean it looks good to me!