Add Whisper Chat Template Support

[Sayan Shaw (sayanshaw24)]

Updates

Neither OpenAI nor HuggingFace has explicit Whisper chat template functionality.

1. OpenAI doesn't expose Whisper via a chat interface like ChatCompletion. Instead, their Whisper API uses raw audio uploads. For finetuning or embedding Whisper in pipelines, they rely on pre-tokenized sequences. They don’t use Jinja2 or chat templates for Whisper at all — it is purely sequence input with prepended tokens like <|startoftranscript|> manually inserted.
2. In HuggingFace transformers, for Whisper, you are expected to pass in the template manually or have it defined in tokenizer_config.json.

However, the expected logic (same for HF and OAI) should emulate concatenating message['content'], with no roles, separators, etc. In this PR, we thereby automatically handle this in ORT Extensions as well.

Validation

• Native C++ testing against baseline output from HuggingFace Whisper model.
• Python unit testing against HuggingFace transformers