Bug in rnorm. (PR#1664)

There appears to be a mild bug, or at least a deficiency, in
rnorm.  The bug becomes apparent when one looks at extremes
of the squares of the values generated by rnorm; rnorm is not
generating quite enough extreme values.

The R version that I am using is 1.4.1; I never got around to installing
1.5.0, and now since 1.5.1 is about to come out ....  However, checking
the 1.5.0 release notes revealed no mention of fixing a bug in rnorm.

After some discussion with colleagues, I replaced the calls to rnorm()
by calls to myrnorm() defined by

	myrnorm <- function(n,mu=0,sigma=1){
		mu + sigma*cos(2*pi*runif(n))*sqrt(-2*log(runif(n)))
	}

which uses the ``(r,theta)'' method of generating random normals.

When I did so, the resulting values were indeed all ``close to'' 0.05,
as they should be.

I also tried the experiment using rchisq(n,1) instead of rnorm(n) (and
then of course taking m = max of x --- rather than max of x^2).  Again
all the resulting values were close to 0.05 as ought to be the case.
(So rchisq() appears to be OK in this regard.)