Add Zipf distribution support for skewed index generation in ModelInput

[Frederick-Zhu]

Summary:
Add Zipf distribution support for generating embedding indices in test utilities.

This change adds an optional zipf_alpha parameter that, when set, uses np.random.zipf() instead of uniform random sampling to
generate indices. This produces a skewed access pattern where some embedding rows are accessed much more frequently than others,
enabling benchmarking scenarios with hot/cold data characteristics.
With alpha ≈ 1.1-1.2, approximately 20% of embedding rows receive ~80% of accesses. This can be useful for testing:

• SSD-offloading strategies
• Caching policies
• Tiered storage performance

Implementation details

• Zipf logic is encapsulated in ModelInput._generate_zipf_indices() static method
• numpy is lazy-imported to avoid breaking backward compatibility when numpy is unavailable
• Falls back to uniform random distribution if numpy import fails
• Optional seed parameter allows controlling numpy's random state independently of PyTorch's seed

Changes

• input_config.py: Add zipf_alpha field to InputConfig dataclass
• model_input.py: Add _generate_zipf_indices() helper; plumb zipf_alpha through generate() and _create_features_lengths_indices()
• test_model.py: Add _generate_zipf_indices() helper; add zipf_alpha support to generate() for both regular and weighted tables
• tests/test_model_input.py: Add unit tests for Zipf distribution functionality
• tests/BUCK: Add test target for test_model_input

Differential Revision: D91909007

[meta-codesync]

@Frederick-Zhu has exported this pull request. If you are a Meta employee, you can view the originating Diff in D91909007.