fix: Fix returned Anthropic reasoning blocks

[jacoblee93]

Preserve structured content blocks returned by Databricks chat models instead of JSON-serializing them.

This fixes Anthropic reasoning responses where the initial streamed chunk contains a structured reasoning block and subsequent chunks contain answer text. Previously, LangChain combined these into a mixed list:

[reasoning_block, "answer text"]

Changes

• Preserve list-based structured content in non-streaming and streaming message conversion.
• After a structured streaming chunk is received, convert subsequent text deltas into indexed text blocks:

  [
      reasoning_block,
      {"type": "text", "text": "answer text", "index": 1},
  ]

[annzhang-db]

generally looks good, can we introduce something like output_version that langchain uses to make this behavior opt-in? just want to avoid breaking anything given the original pr for the string serialization mentions some downstream consumers