Fix thrashing detection and update readme to be correct - #102

The "max heap" column came from `gc_num().max_memory`, which counts only what
the GC charged itself and so misses everything else the process actually paid
for: the mapped sysimage, code, stacks, malloc'd memory, and any heap a GC
mapped without charging. On the bigint/pollard benchmark the two differ by more
than 3x (202 MB accounted vs 643 MB resident).

Record `Sys.maxrss()` in the child instead (getrusage `ru_maxrss`, peak
resident set size), which counts only pages the OS actually backed with
physical memory -- so an allocator that reserves address space it never touches
is not penalised. Reported as "max rss" in the table, the JSON, and
results.csv.

Also size the repeated results.csv columns by the number of runs that produced
results rather than by the requested run count: a run whose child dies is
skipped, and the mismatched column length would then throw instead of reporting
the runs that did succeed.

At 50M points a single run takes roughly 20 minutes, which dominates the whole
`slow` class -- the other benchmark in it, bigint/pidigits, takes 30 seconds.

Default to 10M and read `GCBENCH_RB_TREE_N` for the original scale. What this
benchmark exists to measure is mark performance on a pointer graph whose
minimum linear arrangement is expensive, i.e. mark doing random access with no
prefetching; at 10M points the live set is still over a gigabyte of Point plus
red-black-tree nodes, far beyond any cache or TLB, so that behaviour is intact.