fix: load chat template from chat_template.jinja when available

[ramkrishna2910]

Summary

• Prefer chat_template.jinja file from the model folder over tokenizer_config.json for loading chat templates
• Falls back to tokenizer_config.json if no jinja file is present (preserving existing behavior)
• Matches the Python reference implementation (model_chat.py) which already handles this correctly

Problem

For models like gpt-oss-20b-NPU (and likely other MoE models), OGA's ApplyChatTemplate fails when using the template string from tokenizer_config.json:

[WARNING] OGA chat template failed: Invalid or unsupported chat template.
[WARNING] Using simple fallback template

The fallback template (System: ... User: ... Assistant: ...) produces garbage output because it doesn't match what the model was trained on.

Relates to lemonade-sdk/lemonade#1111

Test plan

• Load gpt-oss-20b-NPU and verify chat completions produce correct output (no fallback warning)
• Load a model without chat_template.jinja (e.g. chatglm3-6b-NPU) and verify it still loads template from tokenizer_config.json
• Load a model with neither file and verify graceful fallback

🤖 Generated with Claude Code

[ramkrishna2910]

Test Results ✅

Tested end-to-end on Windows with gpt-oss-20b-NPU (amd/gpt-oss-20b-onnx-ryzenai-npu).

Key finding: This model's tokenizer_config.json has no chat_template field at all, but ships with a chat_template.jinja file. The old code ignores the jinja file, passes nullptr to OGA, OGA fails, and the fallback template produces garbage.

Before fix (old binary)

• Prompt tokens: 7 (fallback "System:/User:/Assistant:" template)
• Response to "hello": "404\n\``\n\nThis example demonstrates the application of the "404" style..."` — garbage

After fix (PR binary from CI)

• Prompt tokens: 68 (proper jinja template applied)
• Response to "hello": "Hello! 👋 How can I assist you today?" — correct

The token count jump (7 → 68) confirms the model's chat template is now properly wrapping messages.