Support the return_sign argument in logsumexp derivatives

[ump45nose]

scipy.special.logsumexp returns a (log(abs(sum(b * exp(x)))), sign) pair when it is
called with return_sign=True, but the derivatives of logsumexp did not accept that
argument, so both modes raised a TypeError instead of differentiating:

import autograd.scipy.special as asp
from autograd import grad

x = np.array([0.5, 1.5, 2.5])
b = np.array([1.0, -1.0, 1.0])

grad(lambda x: asp.logsumexp(x, b=b, return_sign=True)[0])(x)
# TypeError: make_grad_logsumexp() got an unexpected keyword argument 'return_sign'

This implements the request in #243 (the workaround there was to reimplement logsumexp
in Python and differentiate that).

Accepting the argument is not enough on its own: the returned value is the log of the
absolute value of the sum, so its derivative is

d/dx log(abs(sum(b * exp(x)))) = sign * b * exp(x - ans)

i.e. the sign has to appear in the gradient, otherwise the gradient points the wrong way
whenever the sum is negative. The sign itself is piecewise constant, so only the
log-sum-exp component of the pair has a non-zero derivative (and the tangent of the sign
is zero in forward mode).

Both the VJP (make_grad_logsumexp) and the JVP (fwd_grad_logsumexp) are updated; the
existing code path, where return_sign is not used, is untouched.

Tests

• test_logsumexp_negative_sum pins the sign handling down with a case whose sum is
  exactly negative (b = [1, -3], x = [0, 0], so sign = -1 and ans = log(2)), and
  asserts the gradient equals b * exp(x) / sum(b * exp(x)) as well as checking it
  numerically.
• test_logsumexp_return_sign runs the same shapes/axes/keepdims combinations as the
  existing logsumexp tests with return_sign=True and mixed-sign b, in both modes.

Both fail on main with the TypeError above and pass with this change. The suite is
otherwise unchanged:

$ pytest tests/ -q
634 passed, 13 skipped     # 632 passed, 13 skipped before this change

ruff and ruff-format are clean.

Closes #243