Use epsilon when comparing to 0

[groutr]

This should address an issue found when investigating #2565.

A tolerance needs to be considered in the comparison affine coefficients to 0.
I'm not entirely sure where the best place to test this would be. My initial though was to add a test to test_read.py.

[groutr]

Decided to use transform.array_bounds to compute bounding box in the simple case. array_bounds does not use numpy and therefore doesn't violate the intent of _base.pyx to be numpy-free.
We also use the builtin property to determine if the transform is rectilinear (https://github.com/rasterio/affine/blob/master/affine/__init__.py#L409-L410). This checks for any axis aligned rotation using affine's epsilon.

[groutr]

I'm getting a confusing result when using transform.array_bounds in a test case.

from affine import identity
from rasterio.transform import array_bounds
t = identity.translation(2, 3)
width, height = 10, 10
array_bounds(height, width, t)
# (2, 13.0, 12.0, 3)

I would expect the bounds to be (2, 3, 12, 13). Am I doing something wrong here?

EDIT: It looks like the identity transform is south-up. Any north-up transform involves a reflection (ie. e < 0).