feat: CON-1146: Add more graceful error handling

[OWCramer]

Currently when the worker runs into an issue such as an OOM the process silently dies, the worker goes unhealthy, and the user gets confused. The changes in this PR catch those error and send the error to the user via the standard output. So instead of getting confused and tryign to filter through logs, the request instead comes back with an OOM error.

This approach mirrors implementations from this org's vllm-omni repo and ollama repo.

[velaraptor-runpod]

Can you add this header in the serveless_e2e_test. You can look at this PR, or better just merge that branch here: #338.

Just so we properly test it.

    # Cloudflare in front of rest.runpod.io bans the default Python-urllib UA
    # (HTTP 403, error code 1010, first seen 2026-09-10). A browser UA is not
    # banned; the value just needs to look like one.
    req.add_header(
        "User-Agent",
        "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 "
        "(KHTML, like Gecko) Chrome/126.0 Safari/537.36",
    )

[velaraptor-runpod]

Refer to comment above.

[promptless]

Promptless documentation updates

• Document vLLM worker startup error handling in Serverless troubleshooting updates the Serverless troubleshooting page to document that the vLLM worker now returns a structured startup_error on every job (instead of silently crash-looping) when it hits a recognized fatal startup failure, along with the recognized failure categories and their remediation.