Fix 2D out-of-core feature paths to match the in-RAM path

[darkclad]

Summary

Nyxus computes every feature two ways — an in-RAM calculate() for ROIs that fit
in memory, and a streaming osized_calculate() for oversized ("non-trivial")
ROIs — and the two must return identical values. Exercising the 2D out-of-core
path against its in-RAM twin showed that most of it had silently drifted: 9 of
14 feature families disagreed, two of them crashing the process.

This PR makes every 2D osized_calculate() reuse the proven calculate(), fixes
three pre-existing defects the work surfaced, removes the now-dead streaming code,
and adds equality tests so the invariant can't rot again.

What was broken (2D trivial vs. out-of-core)

Family	Symptom	Cause
GLCM	throws invalid vector subscript	binned with to_grayscale vs bin_intensities
GLRLM, NGTDM	crash 0xC0000005	same, plus NGTDM indexed its matrix by raw grey level
GLSZM, GLDM, GLDZM	wrong / zero values	same binning mismatch
GABOR	garbage (~1e13)	originalScore baseline never assigned → divide by tiny-float
GEOMOMS	shape moments summed intensity (area → 1.9e9)	shared OOC path dropped the intensity-vs-shape intenfunction
CHORDS	shorter chord lengths	OOC subsampled columns instead of scanning all

In every case the in-RAM value was correct and the out-of-core value was wrong.

The fix

Each osized_calculate() now materializes the ROI from the disk-backed pixel
cloud (which these paths already did anyway) and delegates to calculate(), via
two new helpers on LR:

• LR::rebuild_raw_pixels_from_cloud()
• LR::rebuild_aux_image_matrix_from_cloud()

So the two paths agree by construction instead of by duplicated logic.

Pre-existing bugs also fixed

Found while probing degenerate (constant-intensity) ROIs:

• Intensity data loss — constant ROIs replaced INTEGRATED_INTENSITY,
  ENERGY, MODE, ROOT_MEAN_SQUARED and the percentiles with the soft-NAN
  sentinel even though all are well defined; and OOC excess kurtosis used
  KURTOSIS − 3 while in-RAM used Moments4 (differ only on constant data).
• Erosion performance — EROSIONS_2_VANISH ran its erosion chain over
  disk-backed matrices at ~1250× the in-RAM cost (285 s vs 0.23 s at
  100×100); now ~1.1×. It also lacked the in-RAM driver's constant-ROI skip.

Dead code removed

Delegation left the bespoke streaming implementations with no callers. Removed
(~700 lines): GLCMFeature::{Extract_Texture_Features2_NT, calculateCoocMatAtAngle_NT},
GaborFeature::{GaborEnergy_NT2, conv_dud_NT, GetStats_NT},
ZernikeFeature::mb_zernike2D_nontriv, and the whole 2d_geomoments_basic_nt.cpp
translation unit (+ its CMake entry).

Tests

Adds trivial-vs-out-of-core equality mechanics tests in test_ooc_mechanics.py
covering every 2D feature group, on both an ordinary ROI and a degenerate
constant-intensity ROI.

Verification

• C++ gtest: 799/799 pass
• pytest (full dir): 96 pass (7 Arrow failures expected under ALLEXTRAS=OFF)
• trivial == out-of-core: 14/14 families on ordinary and degenerate ROIs
• Zernike OOC vs in-RAM confirmed bit-identical across 10 diverse ROI shapes

Net: roughly −1,950 / +280 lines.

Design note

Delegation materializes the ROI in RAM, so a genuinely huge oversized ROI no
longer streams. This is the right trade here — those paths already built a
disk-backed matrix and were crashing or returning wrong values — and the new
equality tests now guard the invariant, so true streaming can be reintroduced
later without silently diverging again.

[vjaganat90]

This looks fine but can't be merged until PR 389. I suggest, the 2D fixes are real correctness/crash blockers for OOC, they should get CI coverage via pytest, and they are portable to main without #389. As a short rebased/port PR.

As is this PR can't be reviewed or merged.