docs: add canary rollout documentation for RawDeployment mode

## Summary

KServe v0.20 added canary rollout support for InferenceService in Standard (RawDeployment) mode via a new `spec.canary` field. The existing canary docs (`docs/model-serving/predictive-inference/rollout-strategies/canary.md`) explicitly state "Canary rollout strategy is only supported in serverless deployment mode" and use the old `canaryTrafficPercent` field.

## What needs to be documented

- The new `spec.canary` list field with named canary deployments
- `CanarySpec` fields: `trafficPercent` (0-100), `predictor` with required `name`
- Replica scaling: explicit `minReplicas` or derived from `trafficPercent` ratio
- Zero-restart promotion: retarget stable predictor name to canary Deployment
- Rollback by removing canary entry (orphan cleanup)
- Validation rules: Standard mode only, unique names, sum ≤ 100, no autoscaling fields
- `CanaryReady` aggregate condition and per-canary status
- Update existing canary docs to remove "serverless only" limitation note
- Example YAML showing canary definition, promotion, and rollback

## Related PR

- https://github.com/kserve/kserve/pull/5672

## Since

KServe v0.20

Start with docs/model-serving/predictive-inference/rollout-strategies/canary.md and review related PR #5672 for the RawDeployment canary behavior. Document spec.canary, CanarySpec fields, scaling, promotion, rollback, validation, status conditions, and YAML examples. Done means the existing serverless-only note is corrected and the canary lifecycle is fully documented.

Written by the indexing model from the issue text.