fix(ml): cuda_pipeline + trainers infra — CudaSlice throughout

- cuda_pipeline/mod.rs: all Tensor→CudaSlice, DqnGpuData/PpoGpuData GPU-native,
  d2d_subrange helper, build_state_tensor host-interleave + single HtoD
- portfolio_transformer.rs: complete rewrite with GpuLinear + cuBLAS
- trainers/tlob.rs: GpuVarStore + GpuLinear, host-side MSE/MAE
- trainers/tft: StreamTensor trait, MlDevice constructors
- trainers/liquid.rs, mamba2.rs: GpuTensor pairs, f64 loss accumulation
- training/orchestrator.rs: forward_loss(&[f32]) trait interface
- hyperopt/adapters/ppo.rs: MlDevice, PPO::new() constructor
- ml-ppo/lib.rs: fixed cuda_compat→cuda_compile re-export

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/ml-ppo/src/lib.rs

@@ -20,7 +20,7 @@
 #![allow(unsafe_code)] // Required for CUDA kernel launches and cuBLAS FFI
 
 // Re-export shared modules from ml-core for convenience
-pub use ml_core::cuda_compat;
+pub use ml_core::cuda_compile;
 
 pub mod adaptive_entropy;
 pub mod continuous_policy;

crates/ml/src/trainers/tft/model.rs

@@ -2,7 +2,7 @@
 //!
 //! This module provides a trait-based abstraction for the TFT model used by TFTTrainer.
 
-use ml_core::cuda_autograd::GpuTensor;
+use ml_core::cuda_autograd::StreamTensor;
 
 use crate::tft::{TFTConfig, TemporalFusionTransformer};
 use crate::MLError;
@@ -21,11 +21,11 @@ pub trait TFTModel: Send + Sync {
     /// * Quantile predictions [batch, horizon, num_quantiles]
     fn forward(
         &mut self,
-        static_features: &GpuTensor,
-        historical_ts: &GpuTensor,
-        future_ts: &GpuTensor,
+        static_features: &StreamTensor,
+        historical_ts: &StreamTensor,
+        future_ts: &StreamTensor,
         use_checkpointing: bool,
-    ) -> Result<GpuTensor, MLError>;
+    ) -> Result<StreamTensor, MLError>;
 
     /// Get configuration
     fn get_config(&self) -> &TFTConfig;
@@ -39,11 +39,11 @@ pub trait TFTModel: Send + Sync {
 impl TFTModel for TemporalFusionTransformer {
     fn forward(
         &mut self,
-        static_features: &GpuTensor,
-        historical_ts: &GpuTensor,
-        future_ts: &GpuTensor,
+        static_features: &StreamTensor,
+        historical_ts: &StreamTensor,
+        future_ts: &StreamTensor,
         use_checkpointing: bool,
-    ) -> Result<GpuTensor, MLError> {
+    ) -> Result<StreamTensor, MLError> {
         self.forward_with_checkpointing(
             static_features,
             historical_ts,