Don't let partial replication unnecessarily divide push commands

[psalz]

This was surfaced by 2D splits, but can also be constructed (somewhat awkwardly) for 1D: The extra region map lookup into buffer_state.replicated_regions could sometimes cause a box with the same last writer command to be unnecessarily split into several push commands. The fix is to collect all non-replicated boxes and do a final merge before generating the push commands.

[github-actions[bot]]

Check-perf-impact results: (f289611903bd51c6db3f757a138e69d1)

:warning: Significant slowdown (>1.25x) in some microbenchmark results: building command graphs in a dedicated scheduler thread for N nodes - 1 > immediate submission to a scheduler thread / expanding tree topology

Relative execution time per category: (mean of relative medians)

• command-graph : 0.99x
• graph-nodes : 1.00x
• grid : 1.00x
• scheduler : 1.03x
• system : 1.00x
• task-graph : 0.99x

[github-actions[bot]]

Check-perf-impact results: (4c65f1399a47e0eb1340f63004745b17)

:rocket: Significant speedup (<0.80x) in some microbenchmark results: building command graphs in a dedicated scheduler thread for N nodes - 1 > immediate submission to a scheduler thread / expanding tree topology

Relative execution time per category: (mean of relative medians)

• command-graph : 1.02x
• graph-nodes : 1.03x
• grid : 1.00x
• scheduler : 0.99x
• system : 1.01x
• task-graph : 0.98x