Convert hello-world and vllm examples to single-shot mode once single-shot payment lands

[rickstaa]

Context

The request/response examples — hello-world and vllm — are conceptually single-shot (one request in, one response out). But the single-shot payment layer isn't implemented yet (blocked on go-livepeer#3955).

Until it lands, both run on the persistent runner mode (the SDK / static-config default) as a stopgap. On-chain that bills per second of wall-clock for the whole time the session is open, which overbills short request/response calls (a brief job still pays the ~60s preload floor). So these examples are offchain-only in practice for now.

echo is not in scope — realtime trickle streaming is genuinely persistent (Josh set it that way deliberately; cf. his single-shot ping-pong). It stays persistent.

Task (blocked on go-livepeer#3955)

When single-shot on-chain payment is implemented upstream:

• hello-world: register_runner(..., mode="single-shot")
• vllm: runners.json → "mode": "single-shot"
• Re-enable / correct the on-chain (paid) sections for both (drop the persistent-stopgap overbilling caveat)
• Update the main README runner-mode note and the per-example tables

Blocked by

• feat: bill payments for single-shot live runner calls go-livepeer#3955 — single-shot payment layer