Description

This PR implements a reverse keyword-only parameter for the following functions:

• numpy.cumsum
• numpy.cumprod
• numpy.nancumsum
• numpy.nancumprod

When reverse=True, accumulation proceeds right-to-left along the specified axis.
It is equivalent to:

np.flip(fn(np.flip(a, axis=axis), axis=axis), axis=axis)

but implemented using negative-stride views to avoid extra copies.
If out is provided, we accumulate into a flipped view of the output array.

Rationale

• Addresses gh-29144.
• A flip-based approach is efficient, requires no C-API changes, and preserves subclass behavior.
• Scope is limited to high-level APIs (numpy.cumsum, etc.) and does not modify ndarray.cumsum/cumprod or ufunc.accumulate.
  If desired, ufunc.accumulate(reverse=...) can follow in a later PR.

API Changes

• New argument: reverse=False (keyword-only).
• Supported for both regular and NaN-aware variants.
• Works with axis=None, dtype=, and out=.

Examples

>>> import numpy as np
>>> x = np.array([1, 2, 3, 4])
>>> np.cumsum(x)
array([ 1,  3,  6, 10])
>>> np.cumsum(x, reverse=True)
array([10,  9,  7,  4])

>>> np.cumprod(x, reverse=True)
array([24, 24, 12, 4])

>>> y = np.array([1.0, np.nan, 2.0])
>>> np.nancumsum(y, reverse=True)
array([3., 2., 2.])

>>> # Works with multi-dimensional arrays
>>> z = np.arange(6).reshape(2, 3)
>>> np.cumsum(z, axis=1, reverse=True)
array([[ 3,  3,  2],
       [12,  9,  5]])

Tests

• Added new test files covering:
• 1D and multi-D arrays
• axis=None behavior
• dtype overrides
• out parameter
• NaN propagation for nancumsum and nancumprod
• Empty arrays

All new tests pass locally.