🐛 Fix inference pad token id missing bug

[gkumbhat]

Changes

Fix inference.py to work for mistral and support pad_token_id.

Command tested:

_MAX_CONTEXT_LEN=128 _MAX_DECODE_TOKENS=64  _PROMPT_LEN=64  torchrun --nproc_per_node=4 scripts/inference.py --distributed --architecture=hf_pretrained --variant=/models/mistralai/Mistral-Small-3.1-24B-Instruct-2503 --tokenizer=/models/mistralai/Mistral-Small-3.1-24B-Instruct-2503 --device_type=aiu --unfuse_weights --compile --compile_dynamic --compile_dynamic_sendnn --default_dtype=fp16 --fixed_prompt_length=64 --max_new_tokens=64 --timing=per-token --batch_size=1