jgrusewski
20319618b6
refactor(ml-supervised): eliminate candle — GpuTensor/GpuLinear throughout
Zero candle_core/candle_nn imports remain. All 13 source files migrated:
- TFT: GatedResidualNetwork, LSTMEncoder, TemporalAttention → GpuLinear
- Mamba2: SSM layers → GpuLinear + local GpuTensor ops, grad norm host-side
- Liquid CfC: Training rewritten with perturbation-based gradient estimation
- SSD layer: Constructor takes Arc<CudaStream> not VarBuilder
- gpu_tensor.rs: from_candle/to_candle bridge methods deleted
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>