tp: store INTERVAL INTERSECTION's intervals flat per operand (#8004)

Each operand kept a Group per shared group number, indexed directly by
that number. Operands share one GroupTable, so a later operand whose few
keys happen to have high numbers paid 32 bytes for every number below
them: 10 groups could cost 64 MB. Each group also owned its own vector,
one heap allocation per key.

Intervals now arrive grouped and sorted, so each operand keeps its
intervals in one vector in arrival order, with each group a range of it,
and finds a group by binary search over its own groups. A later operand
also skips keys an earlier one lacks, as they cannot make a region.
Memory no longer depends on how keys are numbered.

**Performance**

Instructions per intersection from `perf stat`, and peak memory above
building the tables.

| | before | after |
|---|---|---|
| sched × cpufreq PER cpu, `android_postboot_unlock` | 93M | 93M |
| thread_state × sched PER utid, same trace | 215M | 209M |
| 1M keys × 10 keys, low numbers | 1040M, 183 MB | 595M, 151 MB |
| 1M keys × 10 keys, high numbers | 1080M, 254 MB | 605M, 161 MB |
| 1M keys × 1M keys, 10 shared | 2162M, 422 MB | 1128M, 312 MB |