Introduce inference reserve config for standby instance

[kerthcet]

What would you like to be added:

In serverless scenario, once scaled to 0, and traffic comes, it takes time to recover the service, generally we'll use a standby instance to mitigate the situation, however, usually, a standard llm service may use high-end GPUs which cost a lot.

So we hope to introduce a new config, the inferenceReserveConfig to define the standby instance with lower GPUs, then once traffic comes again, the lower-configured instance will server the traffic immediately, also the response is slow, and trigger the standard-configured instances to roll out.

Why is this needed:

Completion requirements:

This enhancement requires the following artifacts:

• Design doc
• API change
• Docs update

The artifacts should be linked in subsequent comments.

[kerthcet]

/priority important-longterm