[2025.4.1] Gpt-oss multi turn template + add to agentic demo

[dkalinowski]

🛠 Summary

0.35 BFCL multiturn unary

Changes to original chat template:

• introducing reasoning=none - this automaticly adds empty reasoning channel so it forces the model to skip the reasoning part. in reasoning effort placeholder it puts "low", since "none" was not present during gpt-oss training phase
• removed exception when chat history contains both: content and reasoning_content. For some reason, during multi-turn bfcl benchmark i saw model generating both channels, which caused exception during history rendering. instead of exception, the content is rendered and reasoning_content omitted (no idea why i just assumed content might be more insightful)
• replaced thinking with reasoning_content - for some reason openai used thinking field to render reasoning from history. replaced with reasoning_content which is present in requests received during bfcl benchmark.
• fixed issue with parsing empty tool_calls: [] array when rendering chat history. in some cases gpt-oss generates reasoning/content but no tool calls, bfcl sends empty array and chat template accessed index 0 assuming there always is some tool call. new chat templates ignores empty arrays now
• removed |tojson from tool argument rendering. this introduced string escaping which causes escaping in output (chain reaction). openai harmony format assumes no escaping. this was related to both: function call output (result from mcp servers) and function call arguments (input to mcp servers) in chat history

[dkalinowski]

Are you sure there are no functional downsides

No functional downsides visible in simple & multiple datasets

I'm not a fan of creating specific chat template for particular BFCL test category.

I'm not a fan either. But take a look at vllm chat templates: https://github.com/vllm-project/vllm/tree/main/examples - for example mistral. It has specific for parallel calling use case. However, this multiturn chat template should work for all scenarios, therefore I think I will remove original one from PR and just rename multiturn to regular. Are you ok with that? @mzegla