fix: accumulate tool_call_message arguments across streaming chunks

[cpfiffer]

Summary

• Bug: SDKToolCallMessage.toolInput was always {} (empty object) for tool calls streamed from the CLI
• Root cause: The CLI streams tool_call_message in chunks with partial arguments. The SDK's transformMessage() parsed each chunk independently -- JSON.parse("{}") on the first chunk yielded {}, and subsequent chunks with partial JSON either threw or produced { raw: "..." }
• Fix: The background pump now buffers tool_call_message chunks by tool_call_id, concatenating arguments across chunks. The complete message is flushed and transformed when a different message type arrives (e.g., tool_return_message, assistant_message) or the pump ends

Before

{ type: "tool_call", toolName: "web_search", toolInput: {} }

After

{ type: "tool_call", toolName: "web_search", toolInput: { query: "Prime Intellect AI news" } }

Impact

Any SDK consumer that reads toolInput now gets the actual arguments. This was discovered while building tool call display in lettabot -- the display needed to show what arguments tools were called with, but toolInput was always empty.

Test plan

• All 121 existing tests pass (bun test)
• Type-check clean (bun run check)
• Manual test: linked into lettabot via npm link, confirmed toolInput now populated for web_search, conversation_search, Bash, and other tools

Written by Cameron and Letta Code

"The best way to predict the future is to invent it." -- Alan Kay