Improve e2e test isolation, error reporting, and extensibility

[willie-yao]

The e2e tests were recently converted from bash scripts to Go/Ginkgo. They work well but carry over some design choices from the bash era that make them harder to debug, extend, and maintain. This tracks the overall cleanup effort.

Key areas:

• Error messages: Several Eventually calls swallow errors, making failures hard to debug (e.g. pod-not-found shows up as a generic timeout after 120s)
• Test isolation: All 8 test manifests are deployed together in BeforeSuite with a global observedGPUs map tracking state across tests. Tests can't run independently and order matters (test7 is deliberately deployed early)
• Boilerplate: The sharing tests (3, 4, 5, 6) repeat nearly identical verification logic with minor variations
• Per-test driver config: The driver is installed once with fixed Helm values. No test can use different values, and this makes future work like upgrade testing harder

Planned Improvements:

• Make each test self-contained (deploy/cleanup per test context)
• Reduce boilerplate with table-driven tests and shared helpers
• Support per-test Helm values (foundation for upgrade tests)

[willie-yao]

/assign

[nojnhuh]

Feel free to combine any/all of those things into a single PR in separate commits.

[willie-yao]

@nojnhuh I think for supporting per-test helm values, it would be better to have that in a separate PR since it adds a new feature for the tests whereas the other items in the PR are refactoring/cleanup. wdyt?

[nojnhuh]

Landing those things in the same PR won't mess anything up necessarily, so feel free to combine however you want. In general, whatever makes for easiest branch maintenance for you is probably what's easiest to review too.