docs: document graceful pod shutdown (preStop hook and vLLM shutdown-timeout)

## Summary

KServe v0.19 implemented graceful drain for LLM inference pods to prevent in-flight request drops during rolling updates or scale-down events. This involves a preStop hook and vLLM `--shutdown-timeout` flag. There is no documentation for this feature.

## What needs to be documented

- The drain sequence: `preStop sleep (15s) → EPP stops routing → SIGTERM → vLLM drains (≤45s) → pod exits`
- preStop hook (`/bin/sleep 15`) purpose: gives EPP time to stop routing new requests
- vLLM `--shutdown-timeout` flag: computed as `terminationGracePeriodSeconds - 15`, defaults to 45s
- `terminationGracePeriodSeconds` increased from 30s to 60s across all vLLM templates
- How users can override via `spec.template.terminationGracePeriodSeconds`

## Related PRs

- https://github.com/kserve/kserve/pull/5485 (preStop hook + terminationGracePeriod)
- https://github.com/kserve/kserve/pull/5496 (vLLM shutdown-timeout)

## Since

KServe v0.19

Start by reviewing related PRs 5485 and 5496 and the KServe v0.19 vLLM documentation entry points. Document the complete drain sequence, the preStop and shutdown-timeout behavior, the 60-second termination grace default, and the spec.template override so users can verify configuration and expected shutdown timing.

Written by the indexing model from the issue text.