feat: add streaming chunk callback support to generation middleware

[pavelgj]

Add onChunk callback to GenerateParams and withStreamCallback to ModelParams, enabling middleware to observe, transform, or replace streaming chunks. This mirrors the JS SDK's onChunk middleware hook.

• Thread onChunk through GenerateParams and into ModelParams
• Add withStreamCallback to ModelParams for middleware wrapping
• Wire streaming callback through GenerateAction tool loop
• Add StreamingMiddlewareTest covering intercept, transform, and multi-middleware streaming scenarios

[gemini-code-assist]

Code Review

This pull request refactors the streaming implementation to propagate streaming callbacks through the entire middleware chain, enabling middleware to observe or transform response chunks. Key changes include adding onChunk to GenerateParams, streamCallback to ModelParams, and refactoring Genkit.generateStream to delegate to a unified generateInternal method. Comprehensive tests demonstrate middleware capabilities such as re-chunking and content filtering. Feedback suggests moving the supportsStreaming() validation deeper into the call chain to correctly handle middleware-modified model names and eliminate redundant checks.

[gemini-code-assist]

Adding the supportsStreaming() check here ensures that it is performed on the final resolved model, after any middleware has had a chance to modify the model name. This is more robust than checking it at the entry point of generateStream (lines 1587-1590), which is now redundant and can be removed.

          final java.util.function.Consumer<ModelResponseChunk> sc = mparams.getStreamCallback();
          if (sc != null && !model.supportsStreaming()) {
            throw new GenkitException("Model " + modelName + " does not support streaming");
          }

[gemini-code-assist]

The supportsStreaming() check at the beginning of this method (lines 1587-1590) is now redundant if the validation is moved into the model call chain (see suggestion in buildWrappedModelCall). Removing it also avoids a redundant model lookup and ensures that middleware-modified model names are correctly validated.