Add lazy loading for pytorch, tensorflow and the benchmarking module

[deruyter92]

edit: This PR is replaced with #171 (accidentally opened from a fork instead of main repo)

Summary:
This PR fixes import-time failures when optional ML backends are not installed, and makes top-level package imports lighter and safer.

What changed

• Added centralized lazy import helpers in dlclive/utils.py:get_torch(required=..., feature=...), get_tensorflow(required=..., feature=...)
• Updated get_available_backends to use those helpers.
• Enforced clear fail-fast behavior in required runtime paths: dlclive/factory.py now validates backend availability before creating PyTorch/TensorFlow runners.
• Removed eager torch imports from non-required paths: dlclive/benchmark.py, dlclive/live_inference.py
• Made benchmark APIs lazily loaded from package root: dlclive.init now lazily resolves benchmark_videos and download_benchmarking_data.

Why
torch/tensorflow are optional extras. Importing dlclive should not fail if one backend is missing, while backend-specific execution should still raise actionable errors when that backend is actually required.

[deruyter92]

closed in favor of #170

[C-Achard]

(Did you mean #171 ?)