v0.3.0

• Breaking: A regression must also be significant. Each baseline
  comparison carries a reproducible 95% bootstrap confidence interval of its
  difference and its significance, and a change whose interval includes zero
  passes. Metrics with one value per row, such as counts, are judged by their
  threshold alone.
• Breaking: Report JSON changes: every metric records
  isComparedByDefault, statistics count mild and severe outliers, baseline
  comparisons carry their confidence interval and significance, and rows can
  carry allocation sites. Reports and baselines written by 0.2 do not load;
  record them again.
• Breaking: Context switches, page faults, system calls and disk bytes are
  reported but judged only when a threshold or budget names them
  (Benchmark.Metric.isComparedByDefault).
• Breaking: The names of the new built-in metrics are reserved: values a
  workload records under instructions, cycles, contextSwitches,
  pageFaults, systemCalls, diskReadBytes, diskWriteBytes,
  heapAllocations, heapAllocatedBytes or heapNetAllocations are ignored;
  rename custom metrics that use them.
• Breaking: ReportWarning.Kind gains outliers: a compared row warns when
  a tenth of a compared metric's values are severe outliers. Exhaustive
  switch statements over the kind need the new case.
• Timed samples count instructions and cycles, which vary far less with
  machine load than time, and context switches, page faults, system calls and
  disk bytes, each where the platform provides them. The Linux instruction and
  cycle counters are not yet validated on hardware that exposes them.
  BenchmarkRunner(countsProcessEvents:) turns the counting off.
• The package workflow counts every heap allocation, from Swift, C, C++ and
  Objective-C: heapAllocations, heapAllocatedBytes, and
  heapNetAllocations for memory each call keeps. The plugin builds the
  _BenchmarkInterposer library and measures each step in a timing process
  and a counting process that loads it; --interposer passes the library to
  the CLI with your own host. On Linux the counting process also records the
  Swift allocation and reference-counting metrics.
• --allocation-sites records where heap allocations come from: each row
  lists its call stacks with allocations and bytes, in JSON, the console and
  Markdown summaries, and speedscope flame graphs.
• setup: on @Benchmark and the Benchmark initializers makes a fresh input
  for every call outside the timed and counted calls, for plain and Dimension
  benchmarks.
• Dimension curves of cases with three or more sizes carry the growth model
  that best fits them, from O(1) to O(n³).
• --format markdown renders a report summary for pull request comments.
• Batch loops no longer allocate in unoptimized builds, so counts of a debug
  build include only the workload's own allocations.
• The command plugin's package workflow runs on Linux, without Objective-C and
  XCFramework binary targets. Heap counting there needs glibc; elsewhere the
  plugin measures without the counting process.