JENT: Lag preditor health test

In the investigation of issue #21, it became clear that under certain
circumstances (in that case, when the optimizer was turned on) long
repeating sequences could come to dominate the entropy assessment of
the source. In response to this observation, the optimizer was disabled
(in commit 20184e97), which resolved
this issue for the specific platforms where the problem was first
reported. I think this type of failure might reoccur on other
platforms and should be viewed as an observed failure mode. In terms
of SP 800-90B, that suggests that the library should have some sort
of health test that detects this failure mode.

If we examine the entropy estimates produced by the SP 800-90B tests
while the entropy source is in one of these failure modes, we find that
the LRS, t-tuple, and Lag Predictor estimators showed unusually low
assessments. The LRS and t-tuple estimators don't lend themselves to
adaption to an online health test, but the Lag Predictor health test
is amenable to this use.

This patch provides an implementation of a variant of the Lag Predictor
estimator (SP 800-90B Section 6.3.8) as a health test. It includes
cutoffs targeted for the level of significance (alpha) used in the
other health tests. Much like the Lag Prediction Estimator from
SP 800-90B, this health test looks for both global predictability
(successfully predicting more symbols than expected, given an assumed
entropy level of `1/osr`) and local predictability (the presence of
runs of successful predictions longer than would be expected, again
given the assumed entropy level of `1/osr`). Either type of
predictability triggers a health test failure.

As a side observation, increasing the depth of the predictor
(`JENT_LAG_HISTORY_SIZE`) seems to lead to worse (rather than
better) predictions; if this test didn't perturb the noise source
at all, this could not happen. It seems that this may instead be a
consequence of increased  churn in the RAM cache for larger history
depth, so increasing the depth of the predictor has the side effect of
making the system less predictable.

Note: The lag predictor is currently disabled per default by not
defining JENT_HEALTH_LAG_PREDICTOR. The patch is a first merge of the
code into the current tree.

Signed-off-by: Joshua E. Hill
Signed-off-by: Stephan Mueller <smueller@chronox.de>