fix(schema): handle different tool IDs at same index in concatToolCalls

[Ed1s0nZ]

Summary

Some OpenAI-compatible providers (e.g. qwen, deepseek) reuse the same streaming chunk Index for different tool calls, distinguished only by ID. The current concatToolCalls groups chunks solely by Index, causing:

[NodeRunError] cannot concat ToolCalls with different tool id: 'toolu_xxx' 'toolu_yyy'

This PR fixes concatToolCalls to detect when an Index group contains multiple distinct non-empty IDs and splits them into separate sub-groups, each merged independently. Chunks with empty IDs follow streaming order (assigned to the most recently seen non-empty ID). Behavior is unchanged for providers that correctly assign unique indexes per tool call (e.g. OpenAI).

Root Cause

In schema/message.go, concatToolCalls uses map[int][]int keyed by Index only. When provider X sends:

chunk 1: Index=0, ID="tool_a", Name="nmap" 
chunk 2: Index=0, ID="tool_b", Name="sqlmap"

Both land in the same group, and the ID mismatch check fails.

Fix

• Within each Index group, check if multiple distinct non-empty IDs exist
• If yes, split into sub-groups by ID (empty-ID chunks follow the last seen non-empty ID)
• Reassign sequential indexes to all sub-groups
• Existing single-ID-per-index behavior is preserved (no sub-grouping needed)

Test

• Updated different_tool_id test case to verify two tool calls with different IDs at the same index are correctly split into separate merged results
• All existing TestConcatToolCalls sub-tests pass

Fixes #927 #431 #361

[CLAassistant]

All committers have signed the CLA.