reconfigure samplers, new gamma sampler

[simonbyrne]

I've rearranged the samplers, so that they're loaded after the distributions, and take a distribution argument. Also added some more samplers for Gamma.

thoughts?

[coveralls]

Coverage decreased (-0.05%) when pulling 97ae69b on simonbyrne:rgamma into 4c55ac1 on JuliaStats:master.

[simonbyrne]

test results: shape >= 1

BenchmarkTable [unit = mps]
Dist                            |    rmath       MT       GD  
--------------------------------------------------------------
(Gamma(shape=1.5, scale=1.0),)  |  14.9867  20.4117  13.8144  
(Gamma(shape=2.0, scale=1.0),)  |  15.5620  20.0467  14.4154  
(Gamma(shape=3.0, scale=1.0),)  |  15.4051  20.2385  13.2507  
(Gamma(shape=5.0, scale=1.0),)  |  25.2942  20.8716  26.5691  
(Gamma(shape=20.0, scale=1.0),) |  28.5463  21.5892  30.7547  

shape < 1

BenchmarkTable [unit = mps]
Dist                           |    rmath      GS      IP  
-----------------------------------------------------------
(Gamma(shape=0.1, scale=1.0),) |  11.0985  9.0532  8.8145  
(Gamma(shape=0.5, scale=1.0),) |   9.5341  7.1910  9.0511  
(Gamma(shape=0.9, scale=1.0),) |   9.1693  7.3068  9.2258  

I feel that we're close enough to Rmath now that we could safely use these routines now.

[coveralls]

Coverage decreased (-0.05%) when pulling ed7b00d on simonbyrne:rgamma into 4c55ac1 on JuliaStats:master.

[lindahua]

Yes, these samplers already have quite good performance. What we need, however, is to incorporate those in a type-stable way -- if sampler(d::Gamma) is not type stable, we may see a performance hit.