feat(lm): update rig-core v0.31.0, fix reasoning in tool loops, parallel tool calls

[darinkishore]

Summary

• Bump rig-core to v0.31.0, reqwest to 0.13
• Fix reasoning content handling in tool loops: response.choice is OneOrMany<AssistantContent> — providers like Anthropic return [Reasoning, ToolCall] or [Reasoning, Text]. The old code only checked .first(), which caused tool calls behind reasoning blocks to be silently dropped and reasoning blocks to be stripped from chat history (breaking Anthropic's required thinking signature echo)
• Support parallel tool calling: multiple tool_use blocks in a single response are now all executed and their results batched into a single User message
• Extract find_and_execute_tool / execute_tool_batch / push_tool_results helpers to deduplicate the tool-lookup-by-name loop that was copy-pasted 2x
• Add 8 unit tests for classify_choice covering: text-only, single tool call, reasoning+tool call, reasoning+text, reasoning-only fallback, tool-call-wins-over-text, multiple parallel tool calls, image-only fallback

Bugs fixed

1. [Reasoning, ToolCall] responses treated as terminal reasoning — tool call never executed, tool loop exits prematurely
2. Assistant history only included OneOrMany::one(ToolCall), dropping reasoning blocks — subsequent Anthropic API calls would fail due to missing thinking signatures
3. Same .first() bug in the initial call() dispatch

Test plan

• cargo clippy clean (0 warnings)
• cargo test all 30 unit tests + 10 doc-tests pass
• 8 new classify_choice unit tests covering all code paths
• Smoke test: Anthropic claude-sonnet-4 extended thinking + single tool call
• Smoke test: Anthropic claude-sonnet-4 extended thinking + parallel tool calls (3 tools in one response)
• Smoke test: OpenAI gpt-4o-mini tool call (baseline regression check)