fix(gabor): keep filter-response magnitudes real-valued, then vet GABOR vs scikit-image

[darkclad]

The Gabor pipeline stored each filter-response magnitude in a PixIntens (unsigned int) image and thresholded that, so sub-integer responses were floored to 0 -- e.g. several ROIs reported GABOR = 0.0 at the higher frequencies where the true response is small but nonzero. Keep the response real-valued end to end:

• CPU (gabor.cpp): GaborEnergy writes double; calculate() scores from a std::vector
  instead of a PixIntens ImageMatrix.
• non-trivial (gabor_nontriv.cpp): GaborEnergy_NT2 drops the (PixIntens) cast on the magnitude. (Also flags a separate pre-existing bug: osized_calculate leaves originalScore at 0, so the out-of-core baseline score is not wired in.)
• GPU (gabor.cu, cache.h/.cpp, gabor.cpp): the energy montage and host buffers become real-valued -- RealPixIntens (float, 4 bytes like PixIntens, so device/host buffer sizes are unchanged) for the device montage; double for the host scoring buffers. NOTE: the GPU changes are UNVERIFIED here (no CUDA toolchain in this environment) and need a GPU-CI compile/run; CPU and non-trivial paths are unchanged in behavior relative to each other.

Then VET GABOR vs scikit-image (was regression, flagged no-mainstream-oracle):

• Nyxus' Gabor kernel is the canonical Gabor filter: it equals skimage.filters.gabor_kernel (frequency = f0/2pi, sigma_x = sig2lam*2pi/f0, sigma_y = sigma_x/gamma) to ~1e-18.
• An independent reimplementation of the L1-normalize / full-convolve / crop / score pipeline reproduces all 16 golden values (4 DSB2018 ROIs x 4 frequencies) exactly.
• New generator tests/vetting/oracles/gen_gabor_skimage.py (offline; needs scikit-image). gabor_truth updated to the corrected real-valued scores; the gtest is renamed test_gabor_skimage. gabor 0/1 -> 1/1.

gtest 717/717; pytest 72 passed (7 arrow fails are the tiff-only local build; the pytest gabor tests exercise parameter plumbing / CPU-GPU consistency, not feature values). Coverage 604 -> 605 / 758.