Add HF_MODEL to load models directly from huggingface

[yieldthought]

Problem description

Downloading the weights for models is so 2019. We just want to do things like set HF_MODEL=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B and run the demo and have it work first time.

What's changed

Add HF_MODEL as an alternative to setting LLAMA_DIR:

• Loads the model from HuggingFace directly using their organisation/model-name format
• Creates tenstorrent cache tensor files in the existing LLAMA_CACHE_PATH if you set it, otherwise in model_cache/$HF_MODEL
• Tested with mistralai/Mistral-7B-Instruct-v0.3 and works out-of-the-box.
• Work around as_tensor issue to enable models with bias to run again on N150
• Generalise 2d matmul in0_block_w selection
  With these changes even more HF models run out-of-the-box!

Checklist

• All post commit CI passes
• Model regression CI passes (single-card, t3k hanging in main)

[mtairum]

Looks good

[mtairum]

Double check the accuracy numbers, the pipeline was failing due to that.

[mtairum]

Rebased and updated Perf.md.

Re-running:

• single-device demos
• all post commit