Gate the full-length benchmark step on a projection from two shorter ladder steps

For a ladder of benchmark runs at 1M, 2M and full length, where the full run could run out of memory or past a buffer budget, make the driver project the full step before starting it. Take peak RSS (from /usr/bin/time -v) and the tool's own peak-buffer stat at the two earlier steps, and extrapolate linearly to the target length. If the projection is over the limit, or over a fraction (e.g. 0.75) of MemAvailable read at gate time, mark the step SKIP-PROJECTION, write the projection JSON with its basis points, and keep the other lanes running (skip, don't die). Also skip when a basis step did not finish OK or a stat is missing; never project from one point. In a stub test, make the fake program print the buffer stat from an env var, so both an over-limit and an under-limit case can be asserted. My first under-limit case was actually over the limit (100 MiB per 1M x 6.48M = 648 MiB > 512), so do the arithmetic for the test fixtures too.