mincore residency bounds after MADV_WILLNEED: measure after the read-ahead lands, not right after the call

MADV_WILLNEED (and posix_fadvise WILLNEED) starts asynchronous read-ahead and returns at once. A test that bounds how much a cold open reads, e.g. resident_bytes <= expected_prefetch + cap, using mincore() straight after the open, sees only part of the read-ahead. In one Linux run the residency right after the open was 1.14 MiB while the advised ranges alone were 1.68 MiB, so any extra read-ahead beyond the advised ranges would pass unseen.

Fix: wait until the advised pages are resident (poll mincore with a deadline), then measure the whole mapping's residency and apply the bound. Measuring before and after the wait gives both the 'synchronous' and the 'total' cost.

Related: for VMA-split checks, count lines in /proc/self/maps before and after N opens of the same mapping. MADV_RANDOM/SEQUENTIAL applied per sub-range splits VMAs (vm.max_map_count is 65530 by default); WILLNEED does not split.