fix(langchain): keep content blocks of JS input messages on spans

[Harsh23Kashyap]

Fixes #3819

In the JS langchain instrumentors, a message whose content is an array of content blocks reaches the span with only its role. Multimodal messages, tool_result blocks, and any other structured content record as an empty message, because getContentFromMessageData keeps string content and discards everything else.

This change records each content block as a message.contents.N entry, in both openinference-instrumentation-langchain and openinference-instrumentation-langchain-v0, which share the same parsing helper:

• text blocks carry message_content.type = "text" and their text; plain strings inside the array count as text.
• image_url blocks carry message_content.type = "image" and their url, whether the value is a string or an object.
• Unknown block types are skipped, matching the python langchain tracer this mirrors.
• String content, roles, tool calls, and function calls are unchanged.

Test plan:

• New regression test should record content blocks as message contents in test/utils.test.ts of both packages; fails on main and passes with this change.
• pnpm vitest run - 83 passed (langchain), 81 passed (langchain-v0), existing suites unaffected.