This is the beginnings of an attempt to rationalize adapting summary.rq and
boot.rq to use various fitting methods employed by the original fit.

Currently Available fitting methods:

    "br", "fn", "fnc", "sfn" "pfn", "ppro", "conquer", "lasso"

Currently Available Inference methods:

    "rank", "iid", "nid", "ker", "boot", "BLB", "conquer", "extreme"

Currently Available Bootstrap methods:

    "xy", "pwy", "mcmb", "wxy", "wild", "cluster", "BLB", "pxy"

Modifiers for Bootstrap method

    "mofn", "percentile" ...

Problem Space:

    n, p, sparsity

Principles of reorganization:

    o  By default a point estimate using method X, should use method X for
    summary, including "boot", 

    o  Current implementation of fortran for bootstrap all relies on "br"
    except for "cluster", even BLB uses wxy fortran for "br".  If "cluster"
    is used for inference and "sfn" was used for fitting then it is also used
    for inference.

    o  It would be worth exploring whether the fortran wrapper for the
    bootstrap methods is really worthwhile.  E.g.  wouldn't it be better to
    use "fn" in an R loop than "fn" in a fortran loop if the problem were
    large enough?

Preliminary Testing:

    o  test1.R does dense comparison in which "br" does quite well until n
    gets quite large,  i.e. 200,000.  Obviously "sfn" is not so good.

    o  test2.R  adds a factor with 100 levels for which "sfn"  shows some
    better performance.  In both cases the answers agree to 7 digits.

    o  test3.R  first test with multiple taus 

    o  test4.R  second test with multiple taus using fortranization.

    o  test5.R  first test with bootstrap comparison


Exploratory Adventures:

1.  The tuning parameters for preprocessing are still in flux.  For the
bootstrapping strategies, there should be more testing of the choice of m0.
I did a little bit of this:  Optimize on a grid of m0's for various n and p
and then fit a loglinear m0* model in n and p.  I got roughly sqrt(np)
from this and that is what I'm currently using....
