Stop serving stack pods tolerating every taint

[haarchri]

Description of your changes

aicr's bundler stamps a wildcard toleration {operator: Exists} on every pod it renders, Deployments included. On the generated clouds that put cert-manager, the Prometheus stack and the rest on tainted GPU nodes, where they squat on accelerated capacity and their eviction delays autoscaler scale-down.

The generator now rewrites every wildcard the bundle ships: node agents that must reach GPU nodes tolerate exactly the GPU taint the cluster compositions apply, everything else tolerates nothing and schedules on the untainted system pool. Like ALLOW/DROP the tables fail closed, so an aicr bump can't put a new control-plane pod back on GPU nodes. A new unit test pins every cloud, chart values and manifests alike, to keyed tolerations.

The hand-written clouds never stated a wildcard, but one shipped at runtime anyway: the node-exporter chart's default toleration is keyless, and Nebius, Vultr and Existing didn't override it. The unit test can only see stated values, so catching this class needs a rendered cluster. Those three halves now scope node-exporter to the GPU taint, mirroring the generated clouds.

Fixes #

I have:

• Read and followed Modelplane's contribution process.
• Run nix flake check (or ./nix.sh flake check) and made sure it passes.
• Added or updated tests covering any composition function changes.
• Signed off every commit with git commit -s.

[dennis-upbound]

The unanchored patterns are the one fail-open path here.

A renamed upstream workload stops matching the allowlist and the audit goes red, which is the safe direction. But k8s-ephemeral-storage-metrics and kubelet-plugin match anywhere in a pod name, so a future workload that happens to contain either substring gets a wildcard toleration allowed silently — the exact regression this PR exists to catch.

Anchor the ones that can be? The comment says they're unanchored because the generated and hand-written halves name the same workload differently, so if that's only true for some of them, the rest could take ^.

[negz]

The fix LGTM but I'm a bit hesitant to introduce the new E2E test coverage - at least in the same PR.