Studentize loss differentials in SPA and StepM when studentize=True - #871

full-sample standard error. The observed statistic then contains the
sampling error of its estimated standard error, but the bootstrap
distribution does not, so SPA and StepM over-reject. In a Monte Carlo
(AR(1) loss differentials with phi=0.5 and a common factor, a 100-fold
range of model volatilities, 1000 trials, 399 bootstrap replications,
nominal 5%), the consistent p-value rejected a true null in 14.6-18.5% of
trials at T=250, k=20 and 19.3-25.0% at T=100, k=10. With this change it
rejects in 5.6-6.1% and 5.4-6.8% of trials. Not studentizing, the
behavior before bashtage#871, gave 6.6-7.3% and 10.5-12.0%. The ranges cover the
stationary, circular and moving block bootstraps with a block size of 10.

Each replication now computes the standard error of every column from the
bootstrap sample, and divides the re-centered mean by it, as in the
studentized block bootstrap of Gotze and Kunsch (1996). A nested bootstrap
for every replication is far too slow, so the standard error uses a kernel
long-run variance estimator with the kernel implied by the bootstrap:

 * Stationary bootstrap: the weights (1-i/T)(1-p)^i + (i/T)(1-p)^(T-i) of
   Politis and Romano (1994), which give the exact bootstrap variance of
   the mean.
 * Circular block bootstrap: the Bartlett kernel on the circular
   autocovariances, adjusted for the final truncated block, which gives
   the exact bootstrap variance of the mean.
 * Moving block bootstrap: the Bartlett kernel, which equals the
   bootstrap variance of the mean up to edge effects.

A column that is constant in a bootstrap sample has a zero standard
error. This happens when a model differs from the benchmark in very few
periods, for example with 0/1 losses. Dividing by it gave statistics of
order 1e13 in about a third of the replications for a model that differs
from the benchmark in a single period, which made the critical values
huge and removed the power to detect any other model. In those
replications the full-sample standard error is used for that column.

The estimator is computed for all columns with one FFT, so it costs
O(T log T) per column and replication. The full-sample variance used for
the observed statistic and for the consistent recentering threshold uses
the same kernel. Previously it always used the stationary bootstrap
kernel, so the circular and moving block bootstraps give different
variances, and so different p-values, than before. nested=True still
only changes the full-sample variance; the standard errors inside the
bootstrap always use the kernel.

Tests compare the kernels with the exact bootstrap variance of the mean
(the Markov chain of the stationary bootstrap, full enumeration of the
circular block bootstrap, the exact moving block bootstrap variance and
arch.covariance.kernel.Bartlett), the FFT estimator with a direct sum of
autocovariances, and the simulated statistics with a replication-by-
replication recomputation from the same bootstrap draws.