benchmark: add sampling.renorm benchmarks

[xslingcn]

Partial results:

---
top-p renorm
vocab_size: 128512, batch_size: 512, distrib: normal_distribution(std=1), p: 0.1, duration: 2014.94 us, effective bandwidth: 261.24 GB/s
vocab_size: 128512, batch_size: 512, distrib: normal_distribution(std=1), p: 0.9, duration: 2568.39 us, effective bandwidth: 204.95 GB/s
vocab_size: 128512, batch_size: 512, distrib: gumbel_distribution(beta=1), p: 0.1, duration: 2021.48 us, effective bandwidth: 260.40 GB/s
vocab_size: 128512, batch_size: 512, distrib: gumbel_distribution(beta=1), p: 0.9, duration: 2377.93 us, effective bandwidth: 221.36 GB/s
---
top-k renorm
vocab_size: 128512, batch_size: 512, distrib: normal_distribution(std=1), k: 10, duration: 1761.15 us, effective bandwidth: 298.89 GB/s
vocab_size: 128512, batch_size: 512, distrib: normal_distribution(std=1), k: 100, duration: 2299.25 us, effective bandwidth: 228.94 GB/s
vocab_size: 128512, batch_size: 512, distrib: gumbel_distribution(beta=1), k: 10, duration: 1830.94 us, effective bandwidth: 287.49 GB/s
vocab_size: 128512, batch_size: 512, distrib: gumbel_distribution(beta=1), k: 100, duration: 2286.19 us, effective bandwidth: 230.25 GB/s
---
top-k mask logits
vocab_size: 128512, batch_size: 512, distrib: normal_distribution(std=1), k: 10, duration: 1401.01 us, effective bandwidth: 375.72 GB/s
vocab_size: 128512, batch_size: 512, distrib: normal_distribution(std=1), k: 100, duration: 1661.22 us, effective bandwidth: 316.87 GB/s
vocab_size: 128512, batch_size: 512, distrib: gumbel_distribution(beta=1), k: 10, duration: 1551.09 us, effective bandwidth: 339.36 GB/s
vocab_size: 128512, batch_size: 512, distrib: gumbel_distribution(beta=1), k: 100, duration: 1798.39 us, effective bandwidth: 292.70 GB/s

[yzh119]

Let's merge this first, please create another PR for the 2/3 pivot renorm.