Add MKL backends for numpy.fft and scipy.fft and minor fixes

[vchamarthi]

Summary

On the current conda stack, NumPy is the stock conda-forge build, so fft_bench.py measured pocketfft for both numpy.fft and scipy.fft even in an environment with mkl_fft installed. The README environment recipe also resolved to an older oneMKL and mkl_fft generation. This PR adds explicit MKL backend options, records the library versions with every run, and updates the README.

Changes

• fft_bench.py
  • Add --scipy-backend {stock,mkl}. mkl runs the timed region inside scipy.fft.set_backend(mkl_fft.interfaces.scipy_fft). Rows are tagged python-scipy-stock or python-scipy-mkl.
  • Add --numpy-backend {stock,mkl}. mkl routes numpy.fft through mkl_fft.patch_numpy_fft(). Rows are tagged python-numpy-mkl.
  • Both options exit with an error if mkl_fft is missing or the routing does not take effect.
  • Fix the shape check (len(shape) < 0 could never be true).
  • Use numpy.random.default_rng for input generation.
• perf.py
  • Print NumPy, mkl_fft and oneMKL versions in the TAG: lines.
  • Add type hints, use a NamedTuple timer, and move environment printing into print_environment_info().
• fft_bench.c
  • Fix the inverted assertion on the real-input, non-rfft path (ndims != 1 on a path that only runs for 1-D shapes). It was hidden by -DNDEBUG.
  • Add the missing break after case 't'.
• win_compile_all.bat: use CORE-AVX512 instead of COMMON-AVX512.
• README.md
  • The environment recipe now installs mkl_fft and mkl-service with --override-channels, so it resolves to the current oneMKL and mkl_fft releases.
  • Document --numpy-backend and --scipy-backend, with examples.
  • Correct the description of the printed measurements: each measurement is printed, not aggregated.
  • Explain what -c changes in the native benchmark timing.
  • Add a "Comparing platforms" section.

Test plan

Linux (Xeon Gold 6338, socket 0 pinned), before = master 298491f, after = this PR. Python env: Intel channel mkl_fft 2.3.2, MKL 2026.1.0, conda-forge NumPy. Native: icx 2026.1.1, MKL 2026.1.0. Timings are medians of 5 samples over 3 interleaved rounds.

• --numpy-backend mkl / --scipy-backend mkl rejected by master, accepted by this PR
• numpy.fft c128 1024, T=1: master 13.5 us, PR default 13.5 us, --numpy-backend mkl 4.2 us
• scipy.fft c128 1024, T=1: master 10.3 us, PR default 10.3 us, --scipy-backend mkl 6.7 us
• 512x512 c128 T=1 and T=4 with both MKL backends; T=4 is about 1.8x faster than T=1
• Default output: same header and numpy.fft rows; scipy.fft prefix is now python-scipy-stock
• Both options exit with a clear error when mkl_fft is not installed
• TAG: lines report NumPy, mkl_fft and oneMKL versions, None when absent
• Native NDEBUG builds match master (c128 1024, 512x512, f64 rfft 1024 with -c)
• Native assert-enabled build: real 1-D input aborts on master, passes on this PR
• README environment recipe resolves to current mkl_fft and oneMKL; all 8 README commands run as written, including the -P -r example from -P flag appears to be broken in example README command #18
• Native c128 1024, T=1: 18.8 us without -c, 2.7 us with -c
• win_compile_all.bat not tested (no Windows build environment)

[jharlow-intel]

I think approving this sooner rather than later is important, LGTM

[vchamarthi]

I think approving this sooner rather than later is important, LGTM

Fully tested this PR and updated in description.

[jharlow-intel]

super appreciate your thorough testing, lgtm