[antonwolfy]

Follow-up to #396.

_pad_array allocated the zero-padded copy of the input in F order whenever the input had the F-contiguous flag. An input with a single non-unit dimension, such as shape (m, 1), has both the C- and F-contiguous flags, so padding it along the unit axis (n larger than the input length) gave an F-contiguous copy of shape (m, n). That had two effects:

• the result of fft, ifft, rfft and irfft (and of the N-D functions, which pad through the same helper) was F-contiguous, where NumPy returns a C-contiguous array;
• a C-ordered out did not match the layout of the copy passed to MKL, so the result was computed into a temporary array and then copied into out. With perf: check x_arr, not x, when deciding whether irfft can write into out directly #396 this also applies to irfft, which before only got the direct path by checking the original input.

This PR uses F order only when the input is F- but not C-contiguous (PyArray_ISFORTRAN), so a both-contiguous input is padded into a C-ordered copy, as NumPy does. Input that is only F-contiguous is still padded in F order.

Result layout for padding to n=8, NumPy 2.5.4 vs mkl_fft on master:

input shape	axis	NumPy	mkl_fft (master)	mkl_fft (this PR)
(5, 1)	1	C	F	C
(1, 5)	0	C	F	C
(4, 1, 1)	2	C	F	C
(1, 1, 4)	0	C	F	C

The same holds for fft, ifft, rfft and irfft.

Timings for input of shape (4096, 1) padded to n=64 along axis 1 with a C-ordered out, on Lunar Lake, median in ms:

function	master	#396	this PR	#396 + this PR
fft, C-ordered out	2.35	2.34	0.36	0.34
rfft, C-ordered out	0.41	0.41	0.17	0.17
irfft, C-ordered out	0.21	0.52	0.18	0.18
fft, no out	0.39	0.40	0.33	0.33

test_pad_both_contiguous_input checks the result layout and values with and without out for these shapes; it fails on master and passes with this change. The full test suite passes with this PR alone and merged with #396, and the two PRs merge without conflicts.