Prefetch hints over an untrusted offset table: validate the table first, or overlapping entries multiply the cost

A file reader issued madvise(MADV_WILLNEED) on every range listed in a checksummed offset table, to read those ranges ahead before parsing them. The loop ran after the table's checksum check but before the check that entries are contiguous and non-overlapping. A checksum only proves the table is intact, not that it is sane: a crafted file can carry N entries that all cover nearly the whole file. Advice calls then scale as N x file_size / step, which is quadratic in file size because the table grows with the file. Measured on Linux with all pages resident: 8,000 entries in a 297 MB file gave 1.3 TB of advice and a 3.3 s open, against 11 ms for the valid file, before the reader rejected the table. Fix: issue hints in the same walk that enforces the layout rules (expected offset starts at the first segment, stop at the first entry that doesn't start where the previous one ended), so the total is bounded by the file size. Testing tip: when a test counts advised bytes, use a fixture whose size is not a multiple of the page size. A mutant that over-advised by one page went unnoticed on a page-aligned fixture, because the call past the mapping failed with ENOMEM and wasn't counted, and the dropped bytes happened to cancel the extra page.