Add Min-P Sampling Strategy with Tests

[mehulbafnaa]

closes #325

Add Min-P Sampling Strategy (MinPSampling)

This PR introduces a new sampling strategy, Min-P Sampling, based on the recent paper [Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs](https://arxiv.org/abs/2407.01082). Min-P sampling addresses some limitations of existing strategies like Top-k and Top-p, providing improved control over diversity and reducing repetitive token generation.

Motivation

• Enhanced Sampling Quality: Min-P sampling can achieve better diversity and quality compared to existing methods.
• Reduced Repetition: It effectively reduces repetitive outputs common with other sampling methods.

Implementation Highlights

• Adds MinPSampling class under gemma.gm.text.
• Consistent API with existing sampling methods (Greedy, TopkSampling, ToppSampling).
• Minimal invasive changes to existing codebase.

Performance

• Benchmarks confirm comparable decoding speed to Top-p sampling with improved diversity.
• Low overhead; opt-in functionality doesn't affect existing samplers.

Testing

• Unit tests included to verify correct behavior and edge cases.
• Added custom unit tests for additional coverage and robustness.
• All CI checks passing successfully.

Usage Example

from gemma.gm.text import MinPSampling

sampler = MinPSampling(p=0.95)
tokens = sampler.get_next_tokens(logits, rng)

Notes

• This PR intentionally excludes updates to documentation files (README.md and other docs), as requested.

References

• Min-P Sampling paper: [Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs](https://arxiv.org/abs/2407.01082)
• Related Issue: #325

[mehulbafnaa]

@Conchylicultor Please review the PR.