This PR introduces a new subcommand sample2 to address a few issues in the existing sample subcommand:

Problems with existing sample -n

The current implementation uses proportion-based Bernoulli sampling with an amplification factor and early termination. This approach has several limitations:

1. Sampling bias: Records appearing earlier in the input have a higher probability of being selected.
2. Non-deterministic sample size: Does not guarantee an exact target count.
3. Limited control: Two-pass mode (-2) with proportion sampling (-p) doesn't provide explicit number control.

sample2 subcommand solution

-n -2 (Reservoir sampling for fixed-size sampling)

• Provides unbiased, fixed-size sampling with controlled memory usage
• Guarantees exact target count with equal probability for each record
• Memory efficient: tested on large datasets with minimal memory footprint
  • 2,195,354 records: <200 MB memory usage (output: 38GB long read FASTQ)
  • 124,437,023 records: 2.05GB memory usage (output: 43GB short read FASTQ)

-n (Partial in-memory shuffle)

• Produces exact down-sampled size
• Uses complete dataset in memory for unbiased sampling (partitial shuffling)

-p -2 (Two-pass reservoir sampling)

• Two-pass mode now takes effect when -p is specified with -2
• Target sample size is explicitly defined as floor(p × n)
• Combines proportion flexibility with exact count guarantee

-p (Original behavior)

• Retains the original Bernoulli sampling behavior for backward compatibility

Backward compatibility

The original sample subcommand remains unchanged to ensure backward compatibility.