fix a top_heap_words bug in GC statistics

[gasche]

This small change exactly follows @lthls's analysis in #11663, and fixes the repro case in the issue report.

For reference, Vincent's analysis of the bug:

If you look at the code of caml_compute_gc_stats, pool_max_words is very coarsely approximated by taking the sum of the maximums of all the currently domains... but not the orphaned stats.
So you have two domains, one of them terminates and orphans its shared heap, and the program terminates without any other domain adopting the pools. The result is that the reported heap words represent the words from both the main domain and the orphaned pools, while the reported max heap words only takes into account the words from the main domain..
I suspect this can be fixed by adding the orphaned heap stats' max words to the accumulators before looping on the domains.

[abbysmal]

Looks good to me.

Thank you very much for the fix, should I also cherry pick this to 5.0?

[gasche]

should I also cherry pick this to 5.0?

Yes, I think it would be nice.

[abbysmal]

Cherry-picked cc3d9cb to 5.0 at 2f31f0f