Add Intel Gaudi HPU Support

[BartoszBLL]

This PR introduces support for Habana Gaudi (HPU) acceleration and improves device handling across the project. Key changes include:

New HPU-Compatible Dockerfile

• Added Dockerfile.hpu to support Habana Gaudi with PyTorch.
• Uses vault.habana.ai/gaudi-docker as the base image.
• Installs necessary dependencies and sets environment variables for HPU.

Benchmarking on HPU

• Introduced benchmark_hpu.py to compare execution times between CPU and HPU.
• Uses habana_frameworks.torch for HPU acceleration.

Refactored Device Selection

• Unified device selection via get_device() in utils/util.py.
• Replaced multiple scattered torch.device("cuda" if torch.cuda.is_available() else "cpu") checks.
• Now prioritizes HPU if available, then CUDA, falling back to CPU.

Model Loading Adjustments

• Ensured models are first loaded on CPU before transferring to the appropriate device to prevent issues with incompatible state dictionaries.
• Updated synthesizer, vocoder, encoder, and other model scripts accordingly.

Dependency Updates

• Updated requirements.txt to allow NumPy versions >=1.21.0 for better compatibility.

Why

• Adds Habana Gaudi acceleration support for improved training and inference performance.
• Standardizes device management for easier maintainability.
• Enhances model compatibility across different hardware configurations.

Testing

• Verified model loading and inference on HPU, CUDA, and CPU.
• Benchmark (3 runs averaged) shows a 16.25× speedup on HPU vs CPU:
  • HPU: ~23.04 seconds
  • CPU: ~374.27 seconds