Tool-call tests overwrite ToolResult on multi-tool-call first turn

[maragubot]

Description

The can use a tool / can use a tool with args integration tests in all three provider client packages (clients/openai, clients/anthropic, clients/google) accumulate parts from the first ChatComplete turn but overwrite the local result variable on every PartTypeToolCall. When a model emits more than one tool call on the first turn, the tests send back N tool_call_ids in the assistant message but only one matching tool_result, causing the second ChatComplete to fail.

Surfaced by qa during the review of #219 (which fixed #208). The OpenAI variant reproduced ~7% of the time over 15 iterations with GPT-5 Nano:

```
400: An assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'. The following tool_call_ids did not have response messages: call_s442BpDRuyhgmLiEjDi8pPba
```

Anthropic and Google share the same overwrite pattern but didn't reproduce in the qa run.

Affected files

• `clients/openai/chat_complete_test.go` (`can use a tool with args`, ~lines 95-123)
• `clients/anthropic/chat_complete_test.go` (`can use a tool`)
• `clients/google/chat_complete_test.go` (`can use a tool`)

Suggested fix options

• (a) Accumulate `result` into a slice and append every matching tool result to the second-turn message. Most correct; matches real protocol behavior when models emit parallel tool calls.
• (b) Assert exactly one tool call was produced on turn 1, fail-fast otherwise. Simpler but pretends the multi-call case can't happen.

(a) is the right call — the underlying library already supports multiple tool calls, so the tests should too.