Blockwise and Chunked python bindings

[kkiefer]

The module vigra.blockwise module now contains the following functions, all of which are executed in parrallel. For all functions the out parameter is optional unless your source is a hdf5 chunked array, in this case the user is required to specify the out parameter because we can not provide reasonable default names for the associate file and dataset name. If not specify the out parameter will have the same array type / chunked array backend as source.
Unlike the c++ backend, we do not have a way to construct a chunked array containing tinyvector data
on python side as of right now. This implies that one may only call functions on chunked arrays where source as well as out are chunked arrays containing scalar data. In particular this means one can not compute hessianOfGaussianEigenvalues on chunked arrays on python side but calling hessianOfGaussianFirstEigenvalues is perfectly fine because we only need chunked arrays with tinyvector data internally.

1. gaussianSmooth(source, convOptions, out) [CHUNKED ARRAY SUPPORTED]
2. gaussianGradientMagnitude(source, convOptions, out) [CHUNKED ARRAY SUPPORTED]
3. gaussianDivergence(source, convOptions, out)
4. gaussianGradient(source, convOptions, out)
5. laplacianOfGaussian(source, convOptions, out)
6. hessianOfGaussian(source, convOptions, out)
7. hessianOfGaussianEigenvalues(source, convOptions, out)
8. hessianOfGaussianFirstEigenvalue(source, convOptions, out) [CHUNKED ARRAY SUPPORTED]
9. hessianOfGaussianLastEigenvalue(source, convOptions, out) [CHUNKED ARRAY SUPPORTED]
10. symmetricGradient(source, convOptions, out)
11. structureTensor(source, convOptions, out)
12. unionFindWatersheds(source, labelOptions, out) [CHUNKED ARRAY SUPPORTED]
13. labelArray(source, labelOptions, out) [CHUNKED ARRAY SUPPORTED]

Function 1) to 11) return the out paramter.
Function 12) and 13) return a tuple (maxLabel, out).

We support the following datatypes:
Function 1) to 11) -> source.dtype = numpy.float32 out.dtype = numpy.float32
Function 12) and 13) -> source.dtype = numpy.uint8 | uint32 | float32 out.dtype = numpy.float32
Adding additional datatypes is easy to do, please feel free to make an issue or write an e-mail in case u need them.

The option classes [design is not finalized]:
Both option class live in the module vigra.blockwise.

1. BlockwiseLabelOptions.
   -) default constructor
   -) property: blockshape
   -) property: numThreads
   -) property: backgroundValue
   -) property: neighborhood (either "direct" or "indirect")
   -) function: hasBackgroundValue() [returns bool]
   freestanding factory:
   labelOptions(blockshape, neighborhood="direct", backgroundValue=None, numThreads=cpuCount())

2. BlockwiseConvolutionOptionsXD. [X element of {1,2,3,4,5}]
   -) default constructor
   -) property: stdDev
   -) property: innerScale
   -) property: outerScale
   -) property: blockshape
   -) property: numThreads
   freestanding factory:
   convolutionOptions(blockshape, sigma=None, innerScale=None, outerScale=None, numThreads=cpuCount()) [X = len(blockshape)]

[ukoethe]

I cleaned up the commit history and created a new PR #378. Please continue the discussion there.