Improve trt discover and NN interpolation bugfix

[tcollins-hub]

Summary

• Fixed a bug in the nearest-neighbor interpolation CUDA kernel (grid_sample_3d_nearest_kernel) where the per-channel input/output pointers were never advanced across the channel loop. For C > 1, every channel read from input channel 0, and only output channel 0 was ever written — all other output channels were left uninitialized.
• Improved TensorRT/CUDA discovery in CMakeLists.txt (see 21cc043).

Testing

• Added a regression test (testGridSample3dNearestMultiChannel in test.cpp) covering the multi-channel nearest-neighbor bug specifically:
  • Builds a 3-channel, 4×4×4 input with a unique value per (c, d, h, w) voxel and an identity align_corners=true grid, so each output voxel must exactly equal the input voxel at the same coordinates.
  • Seeds the device output buffer with a sentinel value (-1) before running the kernel, so any element the kernel fails to write (as happened for channels > 0 before the fix) is caught rather than masked by leftover memory.
  • Asserts every output element matches the expected value within tolerance.
• Verified the test catches the bug: temporarily reverted the fix and reran — the new test failed via assertion abort as expected. Reapplied the fix and confirmed it passes again.
• Full clean build + test suite: ran a clean cmake -B build configure and build from the committed sources (only pre-existing TensorRT deprecation warnings, no errors), then ctest --output-on-failure — 100% tests passing, including the existing testGridSample3dFloat32 sanity test against reference data in test/data/.