BUG: fix pywt.pad for a zero pad width in 'smooth' and 'antisymmetric' (gh-589)

[Hrafz]

Fixes #589.

A pad width of 0 should leave the array unchanged, but raises in the 'smooth'
and 'antisymmetric' extension modes:

>>> import numpy as np, pywt
>>> x = np.array([1., 2., 3.])
>>> pywt.pad(x, 0, 'smooth')
ValueError: could not broadcast input array from shape (0,) into shape (3,)
>>> pywt.pad(x, (1, 0), 'smooth')
ValueError: could not broadcast input array from shape (0,) into shape (4,)
>>> pywt.pad(x, 0, 'antisymmetric')
ValueError: could not broadcast input array from shape (0,) into shape (3,)
>>> pywt.pad(x, (2, 0), 'antisymmetric')
ValueError: can't extend empty axis 0 using modes other than 'constant' or 'empty'

Six of the nine extension modes return the array untouched for a zero pad
width. 'periodization' is the other exception, and a benign one:
pywt.pad(x, 0, 'periodization') gives [1., 2., 3., 3.], since it promotes
an odd-length axis to even length.

Both failing modes are implemented as numpy.pad callbacks, and both index the
vector from the end using the trailing pad width: vector[-pad_width[1]:] in
pad_smooth and vector[pad_width[0]:-pad_width[-1]] in pad_antisymmetric.
With a trailing width of 0 the first selects the whole vector instead of an
empty slice and the second selects an empty one, hence the two messages above.
The leading width is unaffected because vector[:0] is already empty.

The fix measures both indices from the start of the vector. pad_smooth also
read its right-hand slope with a second negative index (-pad_width[1] - 2),
now expressed relative to the same start-based index. Results are unchanged for
everything that worked before: comparing old against new over 1-D inputs of
length 1 to 5 and both pad widths from 0 to 4, in both modes, the only cases
that differ are ones that raised: a trailing width of 0, plus a length-1 axis
with (0, k) in 'smooth', which raised IndexError.

The message of the pad-width check is also corrected; it read
pad_widths must be > 0 while the check is < 0 and 0 is legal. Happy to drop
that line if you would rather keep the diff to the two callbacks.

Two regression tests are added to pywt/tests/test_dwt_idwt.py:
test_pad_zero_width asserts the no-op property in 1, 2 and 3 dimensions for
every mode except 'periodization', and test_pad_one_sided checks fixed
expected values plus the property that one-sided padding equals the
corresponding slice of two-sided padding, again for every mode except
'periodization'. Both tests fail on main and pass with this change.