Implement true GRU and RNN runtime kernels and layer classes in kernel.cu

[RemindCZ]

Motivation

• GRU and RNN were previously thin placeholders that reused LSTM internals and thus did not implement correct gate math or dimensions.
• Ensure runtime polymorphism so GRULayer/RNNLayer behave correctly when created/initialized through base pointers.

Description

• Added dedicated CUDA kernels kGRUGatesForward, kGRUGatesBackward, kRNNForward, and kRNNBackward implementing GRU (update/reset/candidate) and vanilla RNN (tanh) math.
• Implemented full GRULayer and RNNLayer overrides (Init, InitFromData, Forward, Backward, Update, SaveBest, RestoreBest) that use correct gate dimensions (3 * hidden_size for GRU, 1 * hidden_size for RNN) and integrate with existing cuBLAS/cu kernels.
• Made LSTMLayer methods FreeAll, Init, and InitFromData virtual so derived layers' implementations are invoked through base-class pointers.
• Changes are restricted to kernel.cu and aim only to correct recurrent-layer runtime/gradient behavior and memory layouts.

Testing

• Searched and inspected symbols with rg to verify the new kernels and layer overrides (kGRUGatesForward, kRNNForward, virtual Init), which succeeded.
• Performed static file inspection (nl / code review) to confirm insertion points and gate-dimension adjustments, which succeeded.
• Attempted to compile with nvcc -std=c++17 -c kernel.cu but compilation could not be run in this environment because nvcc is not available; no runtime binary tests were executed.

---

Codex Task