Benchmarks need a grouping axis within a category

[toladnertum]

Every variant so far splits its catalog along two axes, but core only models one of
them (Category), so each variant bends that single axis into a different shape:

Variant	Categories	Groups within a category
VNN-COMP	one (default)	three: test / regular / extended
ARCH-COMP	many (AFF, NLN, AINNCS, …)	one (implicit default)
CORA-COMP	one (contSet)	three: test / sets / sets-batched

The two axes are genuinely different things:

• A category defines the set of benchmarks
• A group partitions this set logically

A first cut is on feature/benchmark-groups: Competition.benchmark_groups() returns
the ordered group names, benchmarks name their group in extra["group"], and the toolkit
submission form renders one subheading per group (default: none, so existing variants
keep a flat list). CORA-COMP uses it today.

What that branch does not do, and what this issue is about:

• the group is a Benchmark.extra convention rather than a field, so nothing validates
  it and nothing can filter or order by it in the database;
• only the toolkit submission form renders it — task details, the benchmark overview and
  the scoreboard still ignore groups (make it similar to how VNN-COMP shows different tracks: | in overview to separate benchmark lists, horizontal dividers on details page, ...);
• VNN's test/regular/extended and ARCH's implicit single group are not expressed in it,
  so the table above is still three different mechanisms rather than one.