fix randnfun lambda

[trefethen]

For large wavelength lambda, randnfun should give results that are nearly constant, but not exactly constant. For example, this is wrong:

>> rng(1), svd(randnfun(100,3))
ans =
   2.187063552007912
   0.000000000000000
   0.000000000000000

I've made a change so that now we get more reasonable results:

>> rng(1), svd(randnfun(100,3))
ans =
   1.033145328179523
   0.067353879945334
   0.000180359044226

The problem was in the "circulant embedding" part of the computation. randnfun uses a subset of an interval on which a periodic random function is constructed. It was defining the subset by

     dom2 = dom(1) + [0 1.2*diff(dom)];

but I've changed this to

    dx = max(0.2,2*lambda/diff(dom));
   dom2 = dom(1) + [0 (1+dx)*diff(dom)];