This fix will preserve subtypes of ndarray when given as input (x)
to the polyval function. In particular, the results for masked
values of a masked array will be masked.

Fixes #2477.

Choices made, and the alternatives rejected (for no particularly strong reason):

1. Show terms in ascending order, to match their internal representation
  * alternative: descending, to match convention
2. Shows 0 terms in gray
  * alternative: omit entirely
  * alternative: show normally to aid comparison
3. Write each term as `basis(ax + b)
  * alternative: write as `basis(u) ... where u = ax + b`
  * alternative: show the normalized polynomial

In future it would perhaps make sense to expose these options to the end user

That function is nose specific and has not worked since `__init__` files
were added to the tests directories.

The `check_roots` function in numpy/polynomial/tests/test_classes.py was
using random numbers without a seed to generate the random roots to be
checked. This made the test sensitive to the random state, with the
result that currently two of the roots are close together and fail the
test to the default seven digit precision when running on the Mac.  The
failure is probably due to a combination of library and compiler
differences between the Mac and the other platforms tested.. The fix
here is to hardwire the test values rather than use random numbers.

The `lbnd` and `scl` parameters of the various polynomial integration
functions  are documented to be scalars. Add tests that those parameters
are checked and ValueError raised when it is not so.