updated phi4 template-update sync-unsloth-bugs

[ankitcoder123]

This pull request introduces the following updates for the Phi‑4 chat template to address issues during fine-tuning and inference:

• User Prompt Formatting:
  Removed the extra assistant prompt token (<|im_start|>assistant<|im_sep|>) from the user formatter. Now, the user messages are formatted as: <|im_start|>user<|im_sep|>{{content}}<|im_end|>
  This change ensures that an assistant prompt is only added when explicitly needed (e.g., via add_generation_prompt=True).

• EOS Token Replacement:
  Ensured that the EOS token is properly replaced with <|im_end|> by setting replace_eos=True in the template registration.

• Dedicated PAD Token:
  Modified the fix_special_tokens method so that if no PAD token is defined, a dedicated pad token (<|dummy_85|>) is assigned rather than reusing the EOS token. This prevents infinite generations during fine-tuning.

• Jinja Template Override:
  Enabled replace_jinja_template=True to guarantee that the custom template is used without default additions.

• Motivation:
  These updates address observed issues in the inference outputs (unexpected tokens and out-of-context responses) when using VLLM for Phi‑4. By refining the chat template and tokenizer special token settings, the model's prompt structure now aligns with the intended design, leading to more accurate and consistent results.

• Changes Made:
  Updated the register_template call in the template registration module.
  Revised the fix_special_tokens function to use <|dummy_85|> as the pad token.
  Tested changes with VLLM to ensure the prompt no longer includes the extra assistant prompt.

• reference links

• https://huggingface.co/microsoft/phi-4/commit/6fbb3d3bbe726c99b4188087b4deeec1bceac5ae#d2h-782009

• https://www.reddit.com/r/MachineLearning/comments/1i23zbo/p_how_i_found_fixed_4_bugs_in_microsofts_phi4/?rdt=63478

• https://unsloth.ai/blog/phi4