refactor: refactor statistics computation for faster stats on large datasets

[vrahier]

Summary

This PR refactors statistics_situations() and statistics() in R/generic_stats.R. They now compute all statistics in one vectorised pass instead of looping over versions and situations. The exported API stays the same.

Motivation

The previous implementation called statistics() once for every version × situation pair. Each of those calls ran:

• one dplyr::summarise() per statistic (around 38 with stat = "all");
• a plyr::join_all() to merge the per-statistic results.

With many situations, this meant tens of thousands of small dplyr calls and joins. Their fixed overhead dominated the execution time.

Changes

• New stats_pairs() builds one long table of observed/simulated pairs (group, situation, Plant, variable, Observed, Simulated) for all versions and situations. format_cropr() now runs once per situation, with all versions stacked together.
• New situation_pairs() holds the per-situation formatting logic that used to live in statistics().
• New compute_stats() computes every requested statistic in a single group_by() + summarise(). Groups and situations keep their order of appearance.
• New stats_description() and check_stat_names() pull the statistic descriptions and the name validation out of statistics().
• statistics() is now a thin wrapper around situation_pairs() and compute_stats().
• cat_situations() / add_situation_col() are no longer used here. all_situations = TRUE now simply relabels the situation column.
• The plyr dependency is removed.