_mi_stats_main.committed.current does goes out of sync with real RSS usage

When tracking `_mi_stats_main.committed.current ` I would expect that it won't go bigger than `VmRSS` in `/proc/../status`.

In practice, I see that for a long running process, I get:
```
comitted_memory:98553561088
used_memory_rss:89664389120
```
which does not make sense to me. We regularly run `mi_stats_merge()` from all the threads to combine the stats into `_mi_stats_main.`

Aggregating data from
`mi_heap_visit_blocks(heap, false, ..)` over all threads provides accurate information about committed memory. It is similar to what I would expect to see (close to used_memory_rss).

Start with _mi_stats_main and mi_stats_merge, then compare their committed-memory accounting with aggregation from mi_heap_visit_blocks across threads. Reproduce the long-running discrepancy against VmRSS and trace which accounting path causes the values to diverge. Done means the reported committed memory is explained and corrected or the intended relationship to VmRSS is documented.

Written by the indexing model from the issue text.