Add Conv3DTranspose, UpSampling3D, and stride-1 fast path for Conv3D

[Dobiasd]

Follow-up to #455. Builds out the rest of the 3D layer family that Keras supports.

Summary

• Conv3DTranspose: dilates the input, then runs a forward conv with strides (1,1,1) and filters reversed along all three spatial axes. New layer class conv_3d_transpose_layer, JSON dispatch entry, and Python exporter (prepare_filter_weights_conv_3d_transpose + show_conv_3d_transpose_layer).
• UpSampling3D: nearest-neighbor resize for volumetric tensors (Keras only supports nearest for the 3D variant). New layer class upsampling_3d_layer, resize3d_nearest helper in tensor.hpp, JSON dispatch, config-only Python exporter.
• Stride-1×1×1 fast path (convolve_accumulative_s1x1x1_3d): collapses the spatial output loops into one big im2col mapping, mirroring convolve_accumulative_s1x1 from the 2D path. Triggers automatically when all strides are 1, so it also speeds up Conv3DTranspose (which always uses internal stride 1).
• Tiny supporting changes: preprocess_convolution_3d gains an optional transposed flag; dilate_tensor_3d gains optional trailing_zeros; generate_filters_3d gains optional transpose; new reverse_size_dim_4_dimension helper.
• Tests: Conv3DTranspose ×5 variants and UpSampling3D ×3 variants added to generate_test_models.py.
• README: Conv3DTranspose added to supported list, Upsampling3D removed from unsupported list, UpSampling1D/2D updated to 1D/2D/3D.

Test plan

• Standalone C++ correctness vs brute-force reference still passes (~1e-6 max diff) with the fast path active for stride-1 cases
• End-to-end Keras → fdeep round-trip passes for all four Conv3DTranspose variants (basic, same, strided 2×2×2, no-bias) and all three UpSampling3D variants (2×2×2, 1×3×5, 2×1×1)
• Conv2D, Conv3D, and the multi-input round-trip from Add Conv3D layer support #455 still pass
• CI runs test_model_exhaustive_test against the updated generate_test_models.py

Out of scope

• SeparableConv3D and DepthwiseConv3D: Keras 3 doesn't expose them (hasattr(keras.layers, ...) == False), so there's nothing to import from
• Further fast-path optimization beyond stride-1×1×1

🤖 Generated with Claude Code