Add Conv3D layer support

[Dobiasd]

Summary

• Implements Conv3D (closes Conv3D #137) using the same im2col + Eigen GEMM strategy as Conv2D, extended with two outer filter-offset loops for the d4 axis
• Supports valid, same, and causal padding, arbitrary strides and dilation rates
• New C++ files: include/fdeep/convolution3d.hpp (math), include/fdeep/layers/conv_3d_layer.hpp (layer wrapper)
• Wires up Conv3D in the JSON dispatch (import_model.hpp) and the Python exporter (convert_model.py)
• Adds Conv3D test cases to generate_test_models.py covering basic / strides+same / dilation+same / no-bias variants

Test plan

• Standalone C++ correctness against a brute-force reference: max diff ~1e-6 across valid/same, strides 1×1×1 / 2×2×2 / 2×2×3 / 1×2×1, and f_d4=1 edge case
• End-to-end Keras → fdeep round-trip: a multi-output Keras model with the four Conv3D variants (basic, strides+same, dilation+same, no-bias) loads and predicts in fdeep with embedded-test diff ~5e-6 vs Keras
• Existing Conv2D unchanged: round-trip diff ~2e-6 in the same multi-input model
• CI runs test_model_exhaustive_test against the updated generate_test_models.py

Out of scope (separate follow-ups)

• Conv3DTranspose, SeparableConv3D, DepthwiseConv3D
• Stride-1 fast path for 3D conv (analogous to convolve_accumulative_s1x1 in 2D)
• UpSampling3D

🤖 Generated with Claude Code