Add raw InferenceService config escape hatch for predictor topology

[tartkri-bot]

Problem

The LLM topology has a powerful escape hatch: the config field on VllmCustomSpec accepts raw YAML that the provider materializes into an LLMInferenceServiceConfig and attaches as a baseRef. This allows users to set any field in the KServe API without waiting for provider updates.

The predictor topology has no equivalent escape hatch. If a user needs a KServe InferenceService feature that the provider doesn't expose (e.g., custom liveness probes, batcher, logger, multi-container predictor, advanced networking), they are completely blocked. They must either:

• Wait for a provider update.
• Create the InferenceService manually via kubectl, bypassing OpenEverest entirely.

Suggested approach

Add a rawConfig or config field to ModelServerCustomSpec (or predictor topology parameters) that accepts a raw InferenceService spec fragment. The provider would:

1. Parse the raw YAML as serving.kserve.io/v1beta1.InferenceServiceSpec.
2. Deep-merge the user-provided spec with the provider-generated spec (user overrides win).
3. Validate that the merged result is still a valid InferenceService.

Alternatively, support a pointer to an externally defined InferenceService or InferenceServiceConfig-like resource.

Why this matters

• KServe's v1beta1 API is large and evolves frequently. The provider will always lag behind.
• Teams with unique requirements need flexibility.
• The LLM topology already proved this pattern works well.

Acceptance criteria

• Predictor topology supports a config or rawConfig field for raw InferenceService YAML.
• Provider deep-merges user config with generated config (structured fields still win).
• Provider validates the merged result and reports parse/validation errors.
• Documentation shows examples of advanced use cases (batcher, logger, custom probes).
• Chainsaw tests verify that raw config overrides are applied correctly.
• LLM topology config behavior is documented for parity reference.