Define VJP and JVP for numpy.flip

[Mohit-Ak]

numpy.flip has no derivative rule registered, so any function that uses it dies at
gradient time:

import autograd.numpy as np
from autograd import grad

grad(lambda x: np.sum(np.flip(x)))(np.array([1.0, 2.0, 3.0]))
# NotImplementedError: VJP of flip wrt argnums (0,) not defined

This is inconsistent rather than intentional: flipud, fliplr and rot90 — the
other members of the same reversal family, sitting on adjacent lines in
numpy_vjps.py/numpy_jvps.py — all have rules. flip is the general-axis
version of exactly those functions, so it just looks like it was missed.

The fix

Two lines, registered next to their siblings:

• Reverse mode: flip is a permutation of the input, so it is its own adjoint —
  flipping the cotangent back along the same axes gives the gradient.
  Passing axis straight through covers all the forms numpy accepts (None,
  an int, or a tuple of ints).
• Forward mode: "same", matching flipud/fliplr/rot90, since flip is a
  linear reindexing.

I registered the JVP as well as the VJP because check_grads exercises both
directions, and leaving forward mode out would have left flip half-broken in
the same way.

Testing

Added test_flip, test_flip_axis and test_flip_axis_tuple to
tests/test_numpy.py, directly alongside the existing test_flipud /
test_fliplr / test_rot90 and following the same check_grads pattern. They
cover the default (all-axes) case, a single named axis, and a tuple of axes on a
3-D array.

All three fail on master with NotImplementedError: JVP of flip wrt argnums (0,) not defined and pass with the change.

Full suite on the branch:

$ python -m pytest tests/ -q
532 passed, 13 skipped, 37 warnings in 3.39s

(The 13 skips are the xarray-dependent tests in test_ufunc_dispatch.py;
xarray isn't installed in my env. Nothing else changed status.)

Linting with the version pinned in .pre-commit-config.yaml (ruff 0.15.22):

$ ruff check autograd/numpy/numpy_vjps.py autograd/numpy/numpy_jvps.py tests/test_numpy.py
All checks passed!
$ ruff format --check <same files>
3 files already formatted

I also checked the gradient values themselves rather than relying on check_grads
alone — comparing against central finite differences for shapes (5,), (4,5)
and (3,4,5) across axis=None/0/1/(0,2) agrees to ~1e-8, and
flip(x, 0) / flip(x, 1) produce gradients identical to the existing
flipud / fliplr rules.

Fixes #725