[Feature] Add tool call accuracy tests for large models to nightly CI

[harvenstar]

Motivation

Tool call tests today only cover a small model (Llama-3.2-1B in test/registered/openai_server/function_call/test_openai_function_calling.py). Large models like DeepSeek V3.2 have no tool call coverage in CI, so bugs like #17593 and #17551 only get caught when users hit them. This issue tracks adding tool call tests to the nightly 8-GPU suite.

Scenarios

These should be common across all models that support tool calling:

Basic

1. Format check — tool_calls is a non-empty list, function.name / function.arguments present, arguments is valid JSON, finish_reason is "tool_calls"
2. Field placement — tool call goes in tool_calls, not content ([Bug] DeepSeek-V3.2 tool calls incorrectly output to content field instead of tool_calls field #17593)
3. Streaming — chunks concatenate to valid JSON, finish_reason correct

tool_choice

4. "required" — always returns tool call
5. "none" — never returns tool call
6. Specific function — returns the specified one

Multi-turn

7. Tool result follow-up — pass tool result back, model replies based on it
8. Thinking + tool call — after tool result, output in content not reasoning_content ([Bug] DeepSeek V3.2: All output marked as reasoning_content after tool call result #17551, DeepSeek specific for now — might be model-internal, will write the test first and see)

Other

9. Parallel tool calls — multiple tool calls in one request
10. Strict mode — strict: true enforces schema

CI integration

Add to test/registered/8-gpu-models/test_deepseek_v32.py, two variants:

Non-MTP:

--tp=8 --dp=8 --enable-dp-attention
--tool-call-parser deepseekv32
--reasoning-parser deepseek-v3

MTP (speculative decoding):

same as above +
--speculative-algorithm=EAGLE
--speculative-num-steps=3
--speculative-eagle-topk=1
--speculative-num-draft-tokens=4
env: SGLANG_ENABLE_SPEC_V2=1

Both in nightly-8-gpu-common.

Plan

Start with DeepSeek V3.2, then extend to GLM / Qwen / others reusing the same scenarios.

Refs

• Small model tests: test/registered/openai_server/function_call/test_openai_function_calling.py
• Parser tests: test/registered/function_call/test_parallel_tool_calls.py
• [Bug] DeepSeek-V3.2 tool calls incorrectly output to content field instead of tool_calls field #17593, [Bug] DeepSeek V3.2: All output marked as reasoning_content after tool call result #17551

[harvenstar]

cc @JustinTong0323 @Kangyan-Zhou