Add onStageComplete

[apsonawane]

This pull request introduces a mechanism for derived classes to hook into the pipeline after a stage completes. The main change is a call to a virtual method that allows subclasses to perform additional actions, such as injecting vision embeddings, after the embeddings stage.

Pipeline extensibility:

• Added a call to OnStageComplete at the end of a pipeline stage in DecoderOnlyPipelineState::RunPipeline, allowing derived classes to inject custom logic (e.g., vision embeddings for Qwen VL) after the embeddings stage.

[Copilot]

Pull request overview

This PR adds a post-stage extensibility hook to the decoder-only pipeline so derived pipeline states can run custom logic immediately after each pipeline stage completes (e.g., injecting vision embeddings after the embeddings stage).

Changes:

• Invoke OnStageComplete(stage_id, next_tokens) at the end of each executed stage inside DecoderOnlyPipelineState::RunPipeline.