provider-kserve adds first-class vLLM engine tuning

The OpenEverest KServe provider (v0.5.0) introduces first-class vLLM engine tuning: context length, GPU memory share, batch size, dtype, quantization, KV cache dtype, and tool-calling and reasoning parsers can all be set through llmEngine.parameters.engine — no hand-written configuration required.

The llm form was also cleaned up: autoscaling, multi-node workers and prefill options now sit behind switches, and parallelism and LoRA settings are grouped together. Autoscaling is usable directly from the UI, which previously always sent a replica count that conflicted with min/max replica settings.

Engine tuning is available now for all KServe llm deployments. To learn more, visit the v0.5.0 release notes.