feat: Support LangChain orchestration client streaming

[KavithaSiva]

Context

Closes SAP/ai-sdk-js-backlog#259.

What this PR does and why it is needed

This PR introduces streaming support for Langchain Orchestration client.

Usage:

const orchestrationConfig: LangchainOrchestrationModuleConfig = {
  llm: {
    model_name: 'gpt-4o'
  },
  templating: {
    template: [
      {
        role: 'user',
        content: 'Write a 100 word explanation about {{?topic}}'
      }
    ]
  }
};

const client = new OrchestrationClient(orchestrationConfig);
const stream =  client.stream([], {
  inputParams: {
    topic: 'SAP Cloud SDK and its capabilities'
  },
  signal: controller.signal
});

for await (const chunk of stream) {
 //Access the streamed chunk here
}

[KavithaSiva]

Adding more tests in a follow-up PR.

[ZhongpinWang]

Suggestions for improvement. Also please do implement tests already in this PR to avoid changes later in case they don't work as expected. 😄

[ZhongpinWang]

I see refactoring needed for the big *_streamResponseChunks function. It would make the code more readable and testable.

[tomfrenken]

Looks like a lot was already considered, still there are some pointers from my side left.

I will continue reviewing the tests.

[tomfrenken]

I would put it as a comment right above this line. Doesn't need to be user-facing, but for us to maintain in the future.

[KavithaSiva]

Some properties are concatenated multiple times, as concat() in AIMessageChunk does not have these properties. We need to custom handle these properties during concatenation. Will handle this in a follow-up BLI: https://github.com/SAP/ai-sdk-js-backlog/issues/322.

I don't think this is a blocker for the streaming feature. What do the others think?

[tomfrenken]

Agreed, but could also be a contribution to LangChain itself (as part of that follow-up ticket), as they hint as concat being the "standard" way to chain chunks together, and it's clearly flawed.

Changes addressed.

[tomfrenken]

A few minor comments, but overall LGTM!

[tomfrenken]

Agreed, but could also be a contribution to LangChain itself (as part of that follow-up ticket), as they hint as concat being the "standard" way to chain chunks together, and it's clearly flawed.