feat: Query ranked/aggregate group compile stats via getattribute

[lgritz]

Post-optimization group metrics active_layers, network_depth, texture_ops, and noise_ops were queryable per group, but the "top N" lists and min/max/median that getstats() prints were computed inline and discarded. Renderers had no way to get them short of scraping the text, and no way to enumerate groups to recompute them.

Add system-level getattribute keys:

stat:compiled_<metric>:top_count    int       # groups with value > 0
stat:compiled_<metric>:top_names    string[n] # ranked, "" padded
stat:compiled_<metric>:top_values   int[n]    # ranked, 0 padded
stat:compiled_<metric>:min|max|median  int    # zero-valued included

Caller's array size picks N; stat:rank_groups still governs only the printed report. Only optimized groups count, and querying never forces optimization.

The gather/sort/aggregate logic moves into gather_group_stats(), which getstats() now also uses, so the report and the queries cannot drift. Keys dispatch before m_mutex is taken, since gather_group_stats() needs only m_all_shader_groups_mutex.

oslexec.h gains documentation only, so there is no API or ABI change.

Testing: testrender gains --print-group-stats, and compstats renders a three-group scene, which testshade cannot do with one group per process. That covers ranking order, the name-ascending tie-break, and exclusion of zero-valued groups from the ranked list while they still count toward min/max/median. The test's filter_re was also dropping the individual ranked entries from both sides of the diff, so the printed top-N was never actually compared; widened to include them.

Assisted-by: Claude Code / Claude Opus 5

[lgritz]

I have no idea why the icx test is failing, but it is certainly not related to the contents of this PR.