feat(ml): integrate predictive AO sequence completion model and Kaggle training pipeline

[Shlok-Parekh09]

Description

This Pull Request introduces the Slope-Domain Sequence Completion AI, which acts as a graceful fallback mechanism for the AO system. When the external frame-quality gate detects partial signal loss (e.g., from mild fog, scattering, or occlusion), this model reconstructs the missing or corrupted wavefront slopes ($\Delta x$, $\Delta y$) per sub-aperture based on the temporal consistency of past frames.

Key additions:

• SlopeCompletionLSTM Model: A mask-conditioned LSTM architecture with a persistence-prior residual head.
• Kaggle Cloud Training Pipeline: Robust dataset extraction, handling of CUDA architecture mismatches, and execution of a massive 100k-sample synthetic dataset training run.
• C Integration: Added predictive_ao_complete_slopes() API for ONNX Runtime integration with zero-dependency static builds.
• Physical Constraints: Integrated Zernike projection and stroke clamping to ensure the model outputs always map into valid actuator space without hardware cross-talk.

Related Issues

Closes #94
(Integrates with the umbrella AI fallback issue #90 and frame-quality gate #93)

Mathematical/Scientific Details

• Architecture: Latent state-space model analogous to a Kalman filter but optimized for Taylor frozen-flow turbulence.
• Loss Formulation: Utilizes a three-component masked slope loss:
  1. Masked reconstruction loss (gradients only flow from missing entries).
  2. Zernike-consistency regularizer (projects slopes onto the modal derivative basis to enforce physical plausibility).
  3. Temporal smoothness regularizer (exploits the Taylor frozen-flow assumption).
• Hardware Constraints: The model output is strictly bounded by a per-spot stroke clamp ($\pm \text{pitch}_{px}/2$) and a Zernike projection that truncates high-order instability via an rcond cutoff.

Verification & Testing Done

Detail the testing performed to verify changes:

• Compiled with GCC / NVCC on Windows/Linux (via CI)
• Run test suite (test_slope_completion.py passed 19/19 checks)
• Check for memory safety (Handled by the existing C static pipeline rules)
• Performance benchmark status (FPS, Latency):
  - Inference Latency: Median ONNX Runtime CPU inference runs at ~1.1 ms.
  - Accuracy: Achieved an RMSE of 0.5201, reflecting a ~30% reduction in error compared to the baseline spatial interpolation (0.7419).

Checklist

• My code follows the code style guides of this project (Linux kernel for C/CUDA, PEP 8 for Python).
• I have updated the documentation (docs/, CONTRIBUTOR.md) and added Doxygen comments to public API functions.
• I have updated CHANGELOG.md with a summary of my changes.

[Shlok-Parekh09]

What about the review anything else to change

[purvanshjoshi]

This is a fantastic contribution. The implementation of the Slope-Domain Sequence Completion AI and the Kaggle training pipeline is well-engineered and thoroughly documented. The inclusion of physical constraints and the C integration architecture are particularly well thought out. The tests demonstrate solid improvements over the baseline. Approved.