Add linear and median filter operations

[johnnynunez]

Summary

Adds NPP linear and rank filtering entry points to FastNPP:

• FilterBoxBorder — mean (box) filter, 8u (C1/C3), 16u (C1), 32f (C1).
• FilterBorder — general 2D convolution with a float coefficient kernel, 32f (C1/C3). Covers Gaussian, sharpen, emboss, arbitrary kernels.
• FilterMedianBorder — per-channel window median, 8u (C1/C3), 16u (C1), 32f (C1).

Each runs as a single fused kernel with NPP_BORDER_REPLICATE semantics.

Dependency

Builds on the FKL filter functors in Libraries-Openly-Fused/FusedKernelLibrary#277 (box + convolution) and #278 (median). Merge those and update the fkl submodule first. Adds CUDA::nppif to the test link line.

Tests

tests/filtering/fastNPP_filter_test.cu cross-checks against nppiFilterBoxBorder, nppiFilterBorder, and nppiFilterMedianBorder (shared aligned pitch, NPP_BORDER_REPLICATE):

[PASS] FilterBoxBorder_8u_C1R    3x3 / 5x5     0 mismatches   (bit-exact)
[PASS] FilterBorder_32f_C1R      3x3 conv      0 mismatches   (max |d| 1.5e-5)
[PASS] FilterMedianBorder_8u_C1R 3x3 / 5x5     0 mismatches   (bit-exact)
ALL PASS

Integrates with CTest. CUDA 13.3 / GCC 11.5, sm_120.

[johnnynunez]

Rebased onto current main, but this PR (linear + median filters (FilterBoxBorder/FilterBorder/FilterMedianBorder)) is BLOCKED: it #includes image_processing/linear_filter.h + median_filter.h, which is not yet merged into FKL main — it lives in FKL #277 + #278 (feat/linear-filter-ops, feat/median-filter).

Verified on RTX PRO 6000 / CUDA 13.3: it does not compile against any merged fkl (4642ba1, a63c000, or 7f25849) because those headers don't exist there yet.

Action: merge the corresponding FKL PR first, then bump the fkl submodule (a follow-up to #22) so these headers are available; this FastNPP PR will then build + pass. I rebased it locally but did not push the rebase to keep CI from going red on an unbuildable branch.

[morousg]

This PR does not compile nor execute. It is using features not present in the FKL

[morousg]

This operations do not exist in the FKL, these are DPPs. We clearly need to work on the instantiable DPPs and allow them to share shared memory tiles