Add truncated autoregressive Wild bootstrap

[benbovy]

It is like the Autoregressive Wild Bootstrap method but autoregression is truncated (i.e., look back) to a given, fixed number of samples.

Unlike the block approach (bootstrap.BlockARWild), it has no "reset" issue but may still greatly optimize the computation using sparse cholesky decomposition (n diagonals sparse matrix -> n samples truncated autoregression). See:

• https://scikit-sparse.readthedocs.io/en/latest/cholmod.html (unmaintained)
• https://github.com/cvxopt/cvxopt
• https://chompack.readthedocs.io/en/latest/documentation.html#numerical-computations