loadbearer v0.5.0

• New sustained-load / thermal soak test. The graded benchmarks are all
  short bursts and measure a machine near its boost clocks; the soak test holds
  every logical CPU under a blended integer + floating-point load for a fixed
  stretch (default 90 s), samples aggregate throughput and CPU frequency once a
  second, and reports the unthrottled peak, the steady-state rate, the
  percentage retained, the throttle onset time, and steady-window
  stability. It is the signal that tells two thin-and-lights with identical
  burst numbers apart. Not scored — measured and shown, like the
  --net-target link probe.
  • loadbearer soak [--duration SECS] [--threads N] [--output FILE] [--json]
    runs it on its own.
  • loadbearer run --soak [--soak-duration SECS] appends it to a full
    assessment; the result is embedded in the result JSON under soak.
  • loadbearer compare shows a SUSTAINED LOAD block — absolute steady
    throughput and retained-vs-own-peak — when every result file carries soak
    data.
• Refinement to memory/bw_read_mt (0.4.0): the per-thread buffers now start
  their timed read together at a barrier (removing staggered-start skew, most
  visible at --duration short) and the per-thread floor rose 16 MiB → 32 MiB
  (insurance against a large shared L3 on server parts). The metric's meaning is
  unchanged; the baseline anchor stays at 28 GiB/s.

Full Changelog: v0.4.0...v0.5.0