Currently, for RNG_TYPE_RND rndsources we apply elaborate uniformity
tests (rngtest) _and_ simple repeated-output tests on samples gathered
serially in global buffers of 512 bytes of output.

In what cases should we gather output of an rndsource into a global
buffer and apply either elaborate uniformity tests or simple
repeated-output tests?

- Not a hardware RNG.  Not uniformly distributed.
  => Example: network packet timings.
  => Elaborate uniformity tests don't make sense and will fail.
  => Small repeated-output tests may raise too many false alarms.
  => Serializing output into a global buffer is needlessly expensive.

- Hardware RNG is per-CPU, output is computed by cryptography.
  => Example: Intel RDRAND/RDSEED.
  => Elaborate uniformity tests should pass.
  => Small repeated-output tests will have adequate statistical power
     to detect plausible failure modes.
  => Unclear whether serializing output for tests into a global buffer
     is helpful or counterproductive.  Current RDRAND/RDSEED seems to
     be per-package, but who knows, maybe per-core might happen.

- Hardware RNG is per-CPU, output is uniform physical observations.
  => Example: ???
  => Elaborate uniformity tests should pass.
  => Small repeated-output tests might have adequate statistical power
     to detect plausible failure modes, but without knowledge of the
     device, hard to say.
  => Unclear whether serializing output for tests into a global buffer
     is helpful or counterproductive.

- Hardware RNG is per-CPU, output is nonuniform physical observations.
  => Example: ???
  => Elaborate uniformity tests will fail, but without knowledge of the
     device or its failure modes, hard to say.
  => Moderate-size repeated-output tests might have adequate
     statistical power to detect plausible failure modes, but without
     knowledge of the device, hard to say.
  => Unclear whether serializing output for tests into a global buffer
     is helpful or counterproductive.

- Hardware RNG is global, output is computed by cryptography.
  => Example: maybe amlogic meson rng, rpi rng, ti/omap rng, ...
  => Elaborate uniformity tests should pass.
  => Small repeated-output tests will have adequate statistical power
     to detect plausible failure modes.
  => Serializing output into a global buffer is reasonable.

- Hardware RNG is global, output is uniform noise.
  => Example: VIA C3 RNG, maybe amlogic meson rng, rpi rng, ti/omap rng, ...
  => Elaborate uniformity tests should pass.
  => Small repeated-output tests _may_ have adequate statistical power
     to detect plausible failure modes, but depends on the device.
  => Serializing output into a global buffer is reasonable.

- Hardware RNG is global, output is nonuniform noise.
  => Example: Geiger counter, maybe allwinner sun8i rng.
  => Elaborate uniformity tests will fail.
  => Small repeated-output tests _may_ have adequate statistical power
     to detect plausible failure modes, but depends on the device.
  => Serializing output into a global buffer is reasonable.