Support partial-dimension FFTs

[michel2323]

The oneMKL FFT wrapper previously rejected any transform region that wasn't all dimensions ("Partial dimension FFT not yet supported"). This was purely a wrapper limitation: the oneMKL DFT descriptor can express these via FWD/BWD_STRIDES, NUMBER_OF_TRANSFORMS, and FWD/BWD_DISTANCE, all of which the support library already exposes.

With this PR, plan_fft/plan_bfft (in-place and out-of-place) and rfft/irfft/brfft accept any contiguous region of dimensions, e.g. fft(A, 1), fft(A, 2), or fft(A, (2, 3)) for a 3D array:

• The four complex plan constructors are unified into _plan_cfft, which programs a region j:k as an R-dimensional transform with explicit column-major strides.
• Dimensions outside the region are batched: trailing dimensions via NUMBER_OF_TRANSFORMS with the block length as distance; leading dimensions as interleaved transforms with distance 1. When batch dimensions exist on both sides (a "middle" region like dim 2 of a 3D array), the trailing side becomes repeated executions at shifted pointer offsets, since a descriptor has only one distance parameter.
• Non-contiguous regions (e.g. (1, 3)) throw an informative error instead of the previous blanket rejection.

Two silently-wrong pre-existing paths are also fixed:

• Multi-dimensional brfft was routed to a non-batched 1D descriptor, so only the first column was transformed and the rest of the output was garbage. It now goes through the complex path.
• The complex-based inverse real FFT (ComplexBasedRealIFFTPlan) zero-padded the missing half of the spectrum instead of reconstructing it by conjugate symmetry, giving inaccurate results. It now performs a proper Hermitian reconstruction (using negative-step range indexing, since Base.reverse(A; dims) falls back to scalar indexing for oneArray).

Tests cover partial regions for even/odd sizes in 2D/3D (complex forward/inverse, in-place, real forward/inverse) against FFTW, plus N-D irfft/brfft which was previously only tested in 1D. The fft testsuite passes 188/188 on an Arc A750.

[codecov]

Codecov Report

❌ Patch coverage is 95.94595% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 80.90%. Comparing base (38876c5) to head (388319a).

Files with missing lines	Patch %	Lines
lib/mkl/fft.jl	95.94%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #615      +/-   ##
==========================================
+ Coverage   79.97%   80.90%   +0.92%     
==========================================
  Files          50       50              
  Lines        3496     3488       -8     
==========================================
+ Hits         2796     2822      +26     
+ Misses        700      666      -34     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.