Fix CLIP inference notebook dependency conflicts on Python 3.12

[fazhangm]

Problem

The hf_pretrained_clip_large_inference_on_inf2.ipynb notebook was failing in CI with two dependency issues:

1. numpy downgrade breaks neuronx-cc: The notebook pinned numpy==1.26.4, but the test environment runs Python 3.12 where neuronx-cc 2.0.254643.0a0 requires numpy>=2.0.0. Installing numpy 1.26.4 caused the Neuron compiler to become incompatible:

   ERROR: neuronx-cc 2.0.254643.0a0+306c5a11 requires numpy>=2.0.0; python_version >= "3.12", but you have numpy 1.26.4 which is incompatible.
   

2. Unpinned transformers pulled breaking major version: Without a version pin, pip resolved transformers-5.8.0 (a major version jump from the 4.x the notebook was written for), which has breaking API changes for CLIP model outputs.

Fix

• Removed numpy==1.26.4 pin — the pre-installed numpy 2.4.4 in the DLC is compatible with all packages including neuronx-cc.
• Pinned transformers==4.57.6 — ensures a compatible 4.x version is used, avoiding breaking changes in transformers 5.x.
• Added %env HF_HUB_DISABLE_PROGRESS_BARS=1 — suppresses xet progress bar rendering errors during non-interactive notebook execution in CI.
• Updated model code to use return_dict=True with output_cpu.logits_per_image[0] — the stable, documented API for accessing CLIP similarity logits (traced Neuron model still returns tuples, accessed via output_neuron[0][0]).

Test plan

• Run notebook on Inf2 with Neuron 2.x DLC (Python 3.12)
• Verify pip install completes without dependency conflicts
• Verify model compilation and inference produce correct top-5 classifications