Major feature release

Major release providing new functionality and laying groundwork for
upcoming functionality.

Breaking changes:

• The HDF5 file format has been changed to allow for non-binary trees
  and to improve performance. It is now both smaller and faster to
  load. However, msprime cannot directly load tree sequence files
  written by older versions. The msp upgrade utility has been
  developed to provide an upgrade path for existing users, so that
  files written by older versions of msprime can be converted to the
  newer format and read by version 0.4.x of msprime.

• The tuples returned by the mutations method contains an element.
  This will break code doing things like

  for pos, node in ts.mutations():
      print(pos, node)
  

  For better forward compatibility, code should use named attributes
  rather than positional access:

  for mutation in ts.mutations():
      print(mutation.position, mutation.node)
  

• Similarly, the undocumented variants method has some major changes:

  1. The returned tuple has two new values, node and index
     in the middle of the tuple (but see the point above about using
     named attributes).
  2. The returned genotypes are by default numpy arrays. To revert
     to the old behaviour of returning Python bytes objects, use the
     as_bytes argument to the variants() method.

New features:

• Historical samples. Using the samples argument to simulate
  users can specify the location and time of all samples explicitly.
• HDF5 file upgrade utility msp upgrade
• Support for non-binary trees in the tree sequence, and relaxation
  of the requirements on input tree sequences using the read_txt()
  function.
• Integration with numpy, with zero-copy access to the low-level C API.
• Documented the variants() method that provides access to the sample
  genotypes as either numpy arrays or Python bytes objects.
• New LdCalculator class that allows very fast calculation of r^2 values.
• Initial support for threading.
• The values returned mutations() method now also contain an index
  attribute. This makes many operations simpler.
• New TreeSequence.get_time() method that returns the time a sample
  was sampled at.

Performance improvements:

• File load times substantially reduced by pre-computing and storing
  traversal indexes.
• O(1) implementation of TreeSequence.get_num_trees()
• Improved control of enabled tree features in TreeSequence.trees()
  method using the leaf_lists and leaf_counts arguments.

Bug fixes:

• Fixed a precision problem with DemographyDebugger. #37
• Segfault on large haplotypes. #29