Bug: when passing unstructured array to structured data type, values get broadcast

[jacanchaplais]

If (N, 4)-dimensional numpy array gets written to HdfEventWriter.pmu, instead of converting the columns to the named fields, each column gets broadcasted over the x, y, z, e elements, resulting in a 4x increase in disk usage.

That is, from this:

array([[-2.52392785e+01,  4.17925601e-02,  1.63704453e+01,
         3.00834579e+01],
       [-1.31029150e+01,  2.16167656e-02,  8.49904537e+00,
         1.56179590e+01],
       [ 4.83393040e+02,  4.55137104e+01,  3.82795685e+02,
         6.18282166e+02],
       ...

what gets stored is this

array([[(-2.52392785e+01, -2.52392785e+01, -2.52392785e+01, -2.52392785e+01),
        ( 4.17925601e-02,  4.17925601e-02,  4.17925601e-02,  4.17925601e-02),
        ( 1.63704453e+01,  1.63704453e+01,  1.63704453e+01,  1.63704453e+01),
        ( 3.00834579e+01,  3.00834579e+01,  3.00834579e+01,  3.00834579e+01)],
       [(-1.31029150e+01, -1.31029150e+01, -1.31029150e+01, -1.31029150e+01),
        ( 2.16167656e-02,  2.16167656e-02,  2.16167656e-02,  2.16167656e-02),
        ( 8.49904537e+00,  8.49904537e+00,  8.49904537e+00,  8.49904537e+00),
        ( 1.56179590e+01,  1.56179590e+01,  1.56179590e+01,  1.56179590e+01)],
       [( 4.83393040e+02,  4.83393040e+02,  4.83393040e+02,  4.83393040e+02),
        ( 4.55137104e+01,  4.55137104e+01,  4.55137104e+01,  4.55137104e+01),
        ( 3.82795685e+02,  3.82795685e+02,  3.82795685e+02,  3.82795685e+02),
        ( 6.18282166e+02,  6.18282166e+02,  6.18282166e+02,  6.18282166e+02)]
        ...

This is an obvious bug! Patch it by converting the (N, 4) array as a view compatible with the HDF5 dataset.