The current implementation of the NIST health tests assumes a
min-entropy estimate of one bit per sample, which is quite conservative.
For so-called "pure" sources (e.g., virtio-random, TPM) it might be nice
to support larger estimates so that the tests catch failed devices more
quickly.
Thus:
- let each pure random source provide an estimate, so that downstreams can override defaults if they want to;
- increase the default estimate for pure sources;
- for pure sources initialize the state machine at source registration time.