Refactor sequence-model engine: validation, early-stopping, global grad-norm clip, sync train, binary IO

[RemindCZ]

Motivation

• Bring the DLL from an LSTM-only demo toward a technically honest sequence-model engine by adding validation support, early stopping, global gradient-norm clipping, proper sync/async training semantics and robust binary state I/O.
• Keep the real LSTM implementation intact while exposing necessary controls (mode, seed, global grad norm) to MQL5 consumers and removing/neutralizing misleading diagnostics.
• Provide a versioned binary format for portable save/load of model topology + weights.

Description

• Added versioned binary header ModelBinHeader and implemented SaveBinaryFile / LoadBinaryFile plus new DLL exports DN_SaveBinaryState and DN_LoadBinaryState that write/read a compact binary representation (magic + version + topology + weights).
• Added validation dataset buffers (X_val, T_val), LoadValidationBatch export and a generalized ComputeDatasetMSE path; added tracking of best_val_loss and a patience-based early stopping flow that snapshots/restores best weights when validation is used.
• Implemented global gradient-norm clipping: new kernel kScaleInplace, helper ApplyGlobalGradNormClipping and integration before optimizer updates; exposed DN_SetGlobalGradNormClip / DN_GetGlobalGradNormClip.
• Fixed training API semantics by adding TrainSync_Locked and wiring DN_Train to synchronous training while preserving the existing StartTrainingAsync_Locked used by DN_TrainAsync; added train/eval mode and RNG seed control with exports DN_SetTrainMode and DN_SetSeed.
• Exposed validation and early-stop controls via exports: DN_LoadValidationBatch, DN_SetEarlyStoppingPatience, DN_GetBestValidationLoss; also neutralized misleading telemetry by returning NaN for unsupported per-layer diagnostics.
• Minor utilities and safety: ComputeDatasetMSE (shared full/validation evaluation), progress updates now optionally track validation metric for best-MSE reporting, and maintenance of early-stopping counters.
• Note: LSTM implementation was kept and remains real/untouched; GRULayer and RNNLayer are still thin derivations of LSTMLayer (placeholders) and therefore do not yet provide a distinct GRU or vanilla RNN algorithm in this change.

Testing

• Performed repository-level automated checks: searched/validated presence of new symbols and exports via rg/nl (symbol checks succeeded); commit created successfully (git commit succeeded).
• Verified file edits and diffs with git diff / git status (changes recorded as one modified file kernel.cu).
• Attempted to run a CUDA toolchain check (nvcc --version) but the environment lacks nvcc so no build/test of CUDA kernels could be performed (no binary compiled or runtime tests executed).
• No runtime functional tests were available in this environment; recommend a CI build (NVCC) and a small training/prediction round-trip (train on toy data, validate save/load, test early-stop) before production use.

---

Codex Task