add grad for det, logdet and inv

[CarloLucibello]

and also trace, diag, diagm, logabsdet, triu, tril
x-ref #45

[CarloLucibello]

merge?

[kirnap]

@CarloLucibello thanks for your efforts. I've just reviewed your implementations. Here are some issues I've discovered, do you have any solution or suggestion to solve these problems:

1. trace Implementation

julia> gradcheck(trace, rand(4,4)) # I am aware of that your test works
julia> my_trace(x) = sum(trace(x)); my_trace(rand(4,4)) # works fine
julia> gradcheck(my_trace, rand(1,1)) # not working

2.logdet Implementation

julia> t1
          4×4 Array{Float64,2}:
          0.81897   0.972307   0.59813   0.408066
          0.736922  0.451799   0.946951  0.141639
          0.127154  0.466463   0.271822  0.872671
          0.864788  0.0628381  0.205276  0.918354

julia> logdet(t1)
	  ERROR: DomainError:
          log will only return a complex result if called with a complex argument. Try log(complex(x)).
          Stacktrace:
          [1] nan_dom_err at ./math.jl:300 [inlined]
          [2] log at ./math.jl:419 [inlined]
          [3] logdet(::Array{Float64,2}) at ./linalg/generic.jl:1261

julia> t2
          4×4 Array{Float64,2}:
          0.177807  0.518924  0.721952  0.305002
          0.244914  0.629782  0.704174  0.394654
          0.686861  0.486215  0.583889  0.902661
          0.756204  0.965346  0.766321  0.209436

julia> logdet(t2)
          -3.283823415533983

julia> logdet(complex(t1))
          -0.7562876304616563 - 3.141592653589793im
# Pls try the following line for a few times, in case want to test by yourself :
julia> [ gradcheck(logdet, eye(3) + rand(3,3)) for i in 1:1000 ];

@denizyuret

1. For all other implementations we need KnetArray Implementations.
2. In the logdet implementation issue we may face the following error:
   julia ERROR: MethodError: no method matching complex(::AutoGrad.Rec{Array{Float64,2}})

[CarloLucibello]

Hi, thanks for review.
Both problems should be fine:

1. trace Implementation

This seems to be bug #46, unrelated to this PR

2.logdet Implementation

This is ok, that error is supposed to manifest when the determinant is negative

[CarloLucibello]

I modified the logdet test to have it more unlikely to fail

[kirnap]

@CarloLucibello I saw your change, other gradient checks are OK.