Align with Python's implementation for handling non-string content of ToolMessage of Ollama client

[riywo]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

So far, ChatOllama doesn't accept if a tool responds non-string content:

    error: "Non string tool message content is not supported\n\nError: Non string tool message content is not supported\n    at convertToolMessageToOllama (file:///Users/ryosuke.iwanaga/.npm/_npx/25982cb890102a2e/node_modules/aethr/node_modules/@langchain/ollama/dist/utils.js:119:15)\n    at file:///Users/ryosuke.iwanaga/.npm/_npx/25982cb890102a2e/node_modules/aethr/node_modules/@langchain/ollama/dist/utils.js:140:20\n    at Array.flatMap (<anonymous>)\n    at convertToOllamaMessages (file:///Users/ryosuke.iwanaga/.npm/_npx/25982cb890102a2e/node_modules/aethr/node_modules/@langchain/ollama/dist/utils.js:129:21)\n    at ChatOllama._streamResponseChunks (file:///Users/ryosuke.iwanaga/.npm/_npx/25982cb890102a2e/node_modules/aethr/node_modules/@langchain/ollama/dist/chat_models.js:730:32)\n    at _streamResponseChunks.next (<anonymous>)\n    at ChatOllama._generateUncached (file:///Users/ryosuke.iwanaga/.npm/_npx/25982cb890102a2e/node_modules/aethr/node_modules/@langchain/core/dist/language_models/chat_models.js:211:34)\n    at process.processTicksAndRejections (node:internal/process/task_queues:105:5)\n    at async ChatOllama.invoke (file:///Users/ryosuke.iwanaga/.npm/_npx/25982cb890102a2e/node_modules/aethr/node_modules/@langchain/core/dist/language_models/chat_models.js:88:24)\n    at async RunnableSequence.invoke (file:///Users/ryosuke.iwanaga/.npm/_npx/25982cb890102a2e/node_modules/aethr/node_modules/@langchain/core/dist/runnables/base.js:1280:27)"

However, Python version handle this case without exception: https://github.com/langchain-ai/langchain/blob/e9e597be8e3b75f3659401d38ddef893d9861ceb/libs/partners/ollama/langchain_ollama/chat_models.py#L542-L582

How about to align the implementation with Python? Or at least just JSON.stringify() for now like ChatOpenAI does?

langchainjs/libs/langchain-openai/src/chat_models.ts

Lines 503 to 506 in b7a9cac

	output:
	typeof toolMessage.content !== "string"
	? JSON.stringify(toolMessage.content)
	: toolMessage.content,

Motivation

This is annoying because some of the models accept the non-string content but some don't. Here is DeepWiki's investigation: https://deepwiki.com/search/relevantcontextthis-query-was_e0a9a7d2-be3b-458c-a716-0d4842a54053

Proposal (If applicable)

First, just JSON.stringify() like OpenAI.

Then, port Python's implementation.

ref. autifyhq/aethr#15