fix: handle parallel tool call streaming with index-based ID tracking

[cpfiffer]

Summary

• Fixes tool call argument accumulation for parallel tool calls (when the agent makes multiple tool calls in a single step)
• The Letta API follows OpenAI's streaming format: only the first chunk per tool call includes tool_call_id, subsequent chunks use index only
• PR fix: accumulate tool_call_message arguments across streaming chunks #55 added accumulation but keyed solely on tool_call_id, so argument chunks (which lack the ID) fell through unbuffered
• Result: N individual tool_call SDK messages with empty toolInput: {} instead of one complete message per call

Changes

• Index-based ID tracking: indexToToolCallId map associates index with tool_call_id from the first chunk, so subsequent chunks resolve correctly
• OpenAI format support: Handles both flat { name, arguments, tool_call_id } and nested { function: { name, arguments } } formats
• Stale queue clearing: Clear streamQueue on each send() to prevent cross-turn message leaks
• stream_event skip: Don't flush pending tool calls when a stream_event arrives (prevents premature flush of incomplete accumulation)

Test plan

• Tested with lettabot on Signal: parallel tool calls (4 simultaneous) now produce 4 complete SDK messages with full arguments
• Before fix: ~57 individual tool_call yields with toolInput: {} for 4 tool calls
• After fix: 4 tool_call yields with fully populated toolInput
• Run existing tool-call-args-accumulation.test.ts suite
• Verify single (non-parallel) tool calls still work correctly

Written by Cameron and Letta Code

"The only way to do great work is to love what you do." -- Steve Jobs