docs: document latency predictor sidecar and LoRA affinity scorer auto-injection

## Summary

KServe v0.20 added two auto-injection features for LLMInferenceService that are not documented:

1. **Latency predictor sidecar**: When the `predicted-latency-producer` plugin is detected in the EPP config, the controller auto-injects training-server and prediction-server sidecar containers via a well-known `LLMInferenceServiceConfig`.

2. **LoRA affinity scorer**: When LoRA adapters are configured, the `lora-affinity-scorer` plugin is automatically injected into the default scheduling profile, scoring endpoints by whether the requested adapter is already active on each pod.

## What needs to be documented

### Latency predictor sidecar
- How to enable: add `predicted-latency-producer` plugin to EPP inline config
- Auto-injection of `kserve-config-llm-scheduler-latency-predictor` well-known config
- What gets injected: training-server and prediction-server sidecar containers
- How to override defaults via custom `LLMInferenceServiceConfig` and `baseRefs`
- Example minimal scheduler config

### LoRA affinity scorer
- Automatic injection when `spec.model.lora.adapters` is configured
- What it does: scores endpoints by active adapter availability
- Injection details: added to `plugins` list, `pluginRef` with `weight: 4` in default scheduling profile
- No manual configuration required — it's fully automatic
- Update existing LoRA adapter documentation to mention this behavior

## Related PRs

- https://github.com/kserve/kserve/pull/5678 (latency predictor sidecar)
- https://github.com/kserve/kserve/pull/5655 (LoRA affinity scorer)

## Since

KServe v0.20

Start with the existing scheduler and LoRA adapter documentation, then review related PRs #5678 and #5655 for the latency predictor sidecar and LoRA affinity scorer behavior. Document the listed enablement, injection, override, and scheduling details, including a minimal scheduler configuration example. Done means both features and the existing LoRA documentation clearly explain their automatic behavior.

Written by the indexing model from the issue text.