Allow reading of arbitrary dimension tipsy aux arrays (e.g. 1D or 3D)

[ibackus]

This addresses issue #379.

This is unfortunately not straightforward, since the tipsy format does not include information about the number of columns in a tipsy array.

There are several different file types, all which have dimensions inferred
using different methods:

• For binary arrays, the number of dimensions are inferred robustly from the
  file size, dtype, and number of particles.
• For gzipped binary arrays, the uncompressed filesize cannot be directly
  read for files >4GB, but in general it is possible to correctly estimate the
  number of dimensions
• For ascii arrays, the number of dimensions are estimated by estimating the
  average size (in bytes) of a single data point and dividing the file size
  by this and the number of particles. This should be fairly robust except in
  some edge cases. A warning is raised if the estimate appears bad.
• Reading of gzipped ASCII arrays has also been implemented, but the number
  of dimensions cannot be estimated

[ibackus]

I'm not sure why this has failed a test. It has failed the python 3.6 test, with the error message:

File "/home/travis/miniconda/envs/3.6_1.13.1/lib/python3.6/site-packages/pynbody-0.41-py3.6-linux-x86_64.egg/pynbody/snapshot/tipsy.py", line 899, in __read_array_from_disk
    array_shape = [n_read, ndim]
UnboundLocalError: local variable 'ndim' referenced before assignment

However, it should not be able to get to that point in the code without ndim being assigned a value. This is especially bizarre, because all the other test builds are passed.