Simplify source classes

[jdbuhler]

This patch cleans up the classes in the sources subdirectory. Changes include

• don't use Numba for the sum in point_source.py -- it was just wrapping a Numpy vectorized sum call, so confers no performance advantage
• simplify call / get_flux methods to sum over an iterator, rather than trying to convert to a list or array first. Rather than call Python sum(), which instantiates a 0 value and adds to it, we grab the first result from the iterator directly, then do in-place adds of the remaining iterator values. This pattern has proved between 3 and 10x faster for our use cases vs using sum(), should never be slower, and should be safe for both bare arrays and Quantity arrays with compatible units.
• for point_source integration, use SciPy's integrate.quad_vec() to do integrals for all values of x in one call, rather than separately for each value
• don't convert iterators to lists unnecessarily where the iterable can be used directly, and update type signatures to match

One point_source integration test was updated to allow the result to differ by a very small amount (1e-15) from the analytically expected 1 as a result of using quad_vec().

It is my contention that the rewritten code is not slower than the original, more complex version. But I would appreciate suggestions on appropriate test cases to convincingly verify this hypothesis.

One data point so far: while testing cosipy with the updated ExtendedSource class using a model (Gassian_on_sphere) with a 2D spatial_shape with len(brightness) = 10, we found that call runs at about the same speed for 1000 spatial points and 3x faster for 10000 points.