Preserve chat message blocks

Treat ChatMessage.blocks (TextBlock, ThinkingBlock, ToolCallBlock, ImageBlock, CacheControl, etc.) as the single source of truth for message content, and be careful not to silently drop information when constructing, streaming, or converting messages. Several bugs stemmed from overwriting the full blocks list with only the new text, which erased previously-parsed tool calls or thinking content:

# Bad: wipes out tool_calls/thinking that chat_from_gemini_response already produced
llama_resp.message.blocks = [TextBlock(text=content)]

# Better: update only the relevant block, or append accumulated text,
# keeping any ToolCallBlock/ThinkingBlock already present
for block in llama_resp.message.blocks:
    if isinstance(block, TextBlock):
        block.text = content
        break

Specific guidance:

• When streaming/accumulating responses, replace or update the specific block type being streamed (text, thinking) without discarding tool-call or other blocks already extracted.
• Prefer deriving tool calls from ToolCallBlock entries in message.blocks rather than duplicating them in additional_kwargs["tool_calls"]; if legacy additional_kwargs handling is kept for backward compatibility, make it a fallback, not a parallel source of truth, to avoid duplicate/conflicting tool calls.
• When merging or splitting content blocks (e.g., for chunking/token splitting), check whether block-specific metadata (like a ThinkingBlock’s provider signature) would be lost; if merge semantics aren’t well understood for a block type, default to not merging/splitting it rather than risk corrupting provider-required data.
• When converting ChatMessage to a provider-specific payload, iterate over message.blocks and handle each block type explicitly (text, image via URL or base64, tool call, cache control) instead of assuming a single string content, so multimodal and tool-call information is not lost in translation.
• When copying or transforming a list of messages (e.g., for XML formatting or memory), copy the whole ChatMessage/block objects rather than just content, so multimodal and other block-specific data survives unchanged.

Following this reduces subtle regressions where tool calls, reasoning/thinking content, or multimodal data disappear during streaming, format conversion, or message transformation.