feat: add CUDA-graph capture for decode forward

- New CudaGraphContext class: warmup -> capture -> replay lifecycle
- One graph per batch_size key, all inputs at fixed workspace addresses
- Added position_ids buffer to InferenceWorkspace (required for graph capture)
- Graph only activates when CUDA backend is the current backend
- Default off (opt-in) due to slight numerical divergence in graph replay
- Sampling stays outside the graph (torch.multinomial uses mutable RNG)
- Resolved circular import: KVCache -> TYPE_CHECKING in attention_backend.py

astrai/inference/core/workspace.py

@@ -88,6 +88,11 @@ class InferenceWorkspace:
             (max_batch_size, 1), dtype=torch.long, device=device
         )
 
+        # Per-step position IDs (must be at a fixed address for CUDA-graph capture).
+        self.position_ids = torch.empty(
+            (max_batch_size,), dtype=torch.long, device=device
+        )
+
         # Split-KV partial-result buffers for decode (persistent, one global
         # alloc per process — mirrors FlashInfer's workspace pattern).
         # Shape: [max_batch_size, max_q_heads, _MAX_SPLITS, head_dim] (o_part)