feat(ml): comprehensive GPU saturation audit — 58 fixes across all 10 models

Phase 1 — Fix broken models (P0):
- Diffusion: wire optimizer_step to actually apply gradients (was no-op)
- TLOB: connect forward pass to projection layers (was Tensor::zeros)
- Mamba2: F64→F32 migration across 5 files (~30x faster on L40S tensor cores)

Phase 2 — Eliminate hot-path GPU sync stalls:
- Mamba2: keep dt on GPU in discretize_ssm (4 functions, no CPU round-trip)
- TFT: gate attention weight logging to eval only (8 syncs/forward eliminated)
- Mamba2: defer loss scalar after backward (pipeline stall removed)
- Mamba2: delete dead gradient clipping (4N wasted GPU syncs removed)

Phase 3 — Enable BF16 for supervised models:
- Flip mixed_precision defaults to true in 4 config locations
- Fix cuda_layer_norm to support BF16/F16 via F32 intermediate

Phase 4 — Raise hyperopt bounds for datacenter GPUs:
- 7 adapters with VRAM-aware tiers (TFT, Liquid, TGGN, KAN, xLSTM,
  Diffusion, TLOB) — L40S gets full hidden_dim range
- Fix L40S tier boundary (was excluded at <48000, now >=40000)

Phase 5 — Update memory estimates:
- 10 param_count estimates updated (DQN 200K→12M, TFT 2M→50M, etc.)
- Fix power-of-two rounding (was wasting up to 49% of budget)
- Correct MODEL_OVERHEAD_MB in DQN/PPO/TFT adapters

Phase 6 — Fix per-epoch CPU bottlenecks:
- PPO: deduplicate double advantage normalization (correctness fix)
- PPO: GPU tensor reward normalization + explained variance
- Fuse per-parameter grad norm to single GPU sync (xLSTM, KAN, TGGN)

Phase 7 — Data pipeline:
- GpuBufferPool: use from_slice (eliminate staging buffer copy)

Phase 8 — Correctness:
- TFT: remove broken .detach() in forward_checkpointed (restore gradients)
- Update stale RTX 3050 Ti doc references

33 files changed, 2451 tests pass, 0 clippy warnings.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

crates/ml/src/diffusion/trainable.rs

@@ -29,6 +29,8 @@ pub struct DiffusionTrainableAdapter {
     learning_rate: f64,
     loss_history: Vec<f64>,
     config: DiffusionConfig,
+    /// Gradients from the last backward pass, consumed by optimizer_step.
+    last_grads: Option<candle_core::backprop::GradStore>,
 }
 
 impl std::fmt::Debug for DiffusionTrainableAdapter {
@@ -84,6 +86,7 @@ impl DiffusionTrainableAdapter {
             learning_rate: lr,
             loss_history: Vec::new(),
             config,
+            last_grads: None,
         })
     }
 
@@ -155,10 +158,18 @@ impl UnifiedTrainable for DiffusionTrainableAdapter {
 
         let loss_val = loss.to_scalar::<f32>().unwrap_or(f32::NAN) as f64;
         self.loss_history.push(loss_val);
+
+        // Store gradients for optimizer_step to consume
+        self.last_grads = Some(grads);
+
         Ok(total_norm.sqrt())
     }
 
     fn optimizer_step(&mut self) -> Result<(), MLError> {
+        if let Some(grads) = self.last_grads.take() {
+            self.optimizer.step(&grads)
+                .map_err(|e| MLError::ModelError(e.to_string()))?;
+        }
         self.step += 1;
         Ok(())
     }