Add Intel® Gaudi® HPU Support

[BartoszBLL]

This PR introduces support for Intel® Gaudi® HPUs, including necessary changes for seamless execution on HPU devices.

Key Changes

• HPU Integration
  • get_device_initial() utility to select the best available device (HPU, CUDA, or CPU).
  • HPU-specific handling in clip.py (wrapping models with Habana optimizations).
• Docker Support
  • New Dockerfile.hpu for building an environment supporting Intel® Gaudi® HPUs.
  • requirements_hpu.txt for HPU-specific dependencies.
• Documentation Updates
  • README.md now includes HPU setup instructions (Docker build, run, and usage).
  • Steps to run benchmarks on HPUs.
• Extended Testing
  • New tests in tests/test_consistency.py for HPU vs. CPU consistency checks.
  • Additional test cases for HPU support.

Benchmark Results

Average across 10 runs

Device	Average Execution Time (s)
HPU	0.4225
CPU	199.0485

How to Test

# Build Docker image for HPU
docker build -t clip_hpu:latest -f Dockerfile.hpu .

# Run the container
docker run -it --runtime=habana clip_hpu:latest

# Inside the container, run tests
pytest

Notes

• Ensure Intel® Gaudi® HPU drivers and runtime are installed.
• Tested with PyTorch 2.4.0 and optimum-habana==1.14.1.
• Dramatic performance gains observed on HPU compared to CPU.