[conv] Fix degenerate spatial extent calculation, grouped channel sizing, and add device wrapper validation tests

[Sadik00789]

Summary

• ClosesConv device wrappers: update() leaves stale geometry and fused-output pointers, two wrappers miss the 2 GiB gate, negative output extents not rejected #3545

This PR resolves stale geometry/pointer state across convolution device wrappers, adds missing 2 GiB overflow gates, clamps degenerate spatial extents, fixes 3D grouped convolution channel sizing, and adds unit test coverage:

1. Device Wrapper State & Geometry Refresh: Fixed update() in convolution device wrappers (direct_convolution.h, implicit_gemm_convolution.h, implicit_gemm_convolution_fusion.h) to properly refresh internal tensor geometry and fused-output pointers rather than retaining stale parameters.
2. Degenerate Spatial Extents & Gating:
   • Clamped output spatial dimensions (P, Q, Z) to 0 in conv2d_problem_size.h and conv3d_problem_size.h when (Dim + pads - filter_span) < 0.
   • Added strictly positive spatial extent validation in can_implement() across SM90/SM100 collective and kernel layers.
3. 2 GiB Tensor Offset Overflow Gating: Enforced 2 GiB overflow validation across wrappers missing tensor offset bounds checks.
4. 3D Grouped Convolution Sizing: Updated filter_c() and filter_size() in conv3d_problem_size.h to divide channels by groups (C / groups).
5. CuTe Header Dependencies: Included <cute/tensor.hpp> in convnd_problem_shape.hpp so CuTe utilities (take, reverse, insert, for_each, make_seq) resolve across standalone translation units.
6. Validation Test Coverage: Added test/unit/conv/device/conv_validation_and_update.cu to verify degenerate spatial extent clamping, 2 GiB gating, and runtime pointer/geometry updates.

---

Files Changed

• include/cutlass/conv/conv2d_problem_size.h: Clamped P and Q spatial extents to 0 against negative numerators.
• include/cutlass/conv/conv3d_problem_size.h: Clamped Z, P, Q spatial extents and fixed filter_c() / filter_size() channel sizing for grouped convolutions.
• include/cutlass/conv/convnd_problem_shape.hpp: Added #include "cute/tensor.hpp" for CuTe tuple algorithm definitions.
• include/cutlass/conv/device/direct_convolution.h: Updated update() geometry refreshes and runtime parameter validation.
• include/cutlass/conv/device/implicit_gemm_convolution.h: Added 2 GiB overflow checks and fixed stale pointer caching on update().
• include/cutlass/conv/device/implicit_gemm_convolution_fusion.h: Fixed stale fused-output pointer updates and refreshed geometry state.
• include/cutlass/conv/collective/sm90_implicit_gemm_gmma_ss_warpspecialized.hpp: Added spatial extent validity checks in can_implement().
• include/cutlass/conv/kernel/sm100_implicit_gemm_tma_warpspecialized.hpp: Added problem shape validation in kernel execution entry points.
• test/unit/conv/device/CMakeLists.txt: Registered cutlass_test_unit_conv_device_validation test target.
• test/unit/conv/device/conv_validation_and_update.cu: Added test suite for degenerate extents, 2 GiB offset gating, and wrapper update() refreshes.

---

Testing

Built and verified passing with GCC 13 and CUDA 12.6 targeting SM86:

ninja cutlass_test_unit_conv_device_validation
./build/test/unit/conv/device/cutlass_test_unit_conv_device_validation
[==========] Running 3 tests from 1 test suite.
[----------] Global test environment set-up.
[----------] 3 tests from ConvDeviceValidation
[ RUN ] ConvDeviceValidation.DegenerateSpatialExtentsNotNegative
[ OK ] ConvDeviceValidation.DegenerateSpatialExtentsNotNegative (0 ms)
[ RUN ] ConvDeviceValidation.TwoGiBOverflowGate
[ OK ] ConvDeviceValidation.TwoGiBOverflowGate (0 ms)
[ RUN ] ConvDeviceValidation.UpdateGeometryAndPointerRefresh
[ OK ] ConvDeviceValidation.UpdateGeometryAndPointerRefresh (0 ms)
[----------] 3 tests from ConvDeviceValidation (0 ms total)
[----------] Global test environment tear-down
[==========] 3 tests from 1 test suite ran. (0 ms total)
[ PASSED ] 3 tests.