Add memory-efficient InfiniteHashSampler and infinite sampler tests

[gertln]

PhysicsNeMo Pull Request

Description

Introduces InfiniteHashSampler, a new memory-efficient infinite sampler designed for very large datasets (billion+ samples) that uses hash-based randomization without storing full index arrays.
Tests for both infinite samplers have been added.

• Hash-Based Randomization: Deterministic pseudo-random sampling using efficient hash function
• Distributed Training Support: Full compatibility with DistributedDataParallel (DDP)
• Billion-Scale Ready: Tested with datasets up to 10 billion samples
• Sequential Fallback: Option to disable randomization for sequential access

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.
• The CHANGELOG.md is up to date with these changes.
• An issue is linked to this pull request.

Dependencies