Compare with astropy also with log model

The test previously failed just because the approximate Lomb-Scargle
periodogram had values slightly above 1 (both in LombScargle.jl and in
Astropy), so log(1 - p) gave a DomainError.  Change the frequency grid
and increase the number of input points in order to fix this issue.

test/astropy.jl

@@ -3,7 +3,7 @@
 using LombScargle, Base.Test, PyCall
 PyCall.@pyimport astropy.stats as ast
 
-ntimes = 201
+ntimes = 401
 t = linspace(0.01, 10pi, ntimes)
 t += step(t)*rand(ntimes)
 for f in (x -> sinpi(x), x -> sin(x) + 1.5*cospi(4*x) + 3)
@@ -30,15 +30,17 @@ for f in (x -> sinpi(x), x -> sin(x) + 1.5*cospi(4*x) + 3)
     s = f(t)
     # "psd" normalization in LombScargle slightly differ from that of
     # Astropy and the test would fail if we includ it.
-    for fitmean in (true, false), nrm in ("standard", "model")
+    for fitmean in (true, false), nrm in ("standard", "model", "log")
         f_jl, p_jl = freqpower(lombscargle(t, s, fit_mean = fitmean,
                                            normalization=nrm,
-                                           maximum_frequency=20))
+                                           maximum_frequency=10,
+                                           samples_per_peak=10))
         f_py, p_py =
             ast.LombScargle(t, s,
                             fit_mean = fitmean)[:autopower](method="fast",
                                                             normalization=nrm,
-                                                            maximum_frequency=20)
+                                                            maximum_frequency=10,
+                                                            samples_per_peak=10)
         @test_approx_eq f_jl f_py
         @test_approx_eq p_jl p_py
     end