phase 5: naive multi-head attention

- Batched GEMM via cublasGemmStridedBatchedEx
- Causal mask CUDA kernel (F32 + BF16)
- Element-wise scale CUDA kernel (F32 + BF16)
- attention() composing: batched_matmul + scale + causal_mask + softmax
- Fixed to_device/contiguous infinite recursion (GPU contiguous via CPU round-trip)
- 5 attention tests passing (max_err < 3e-7 F32)
- Total: 61 tests passing across all crates

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

crates/xserv-tensor/src/tensor.rs

@@ -137,8 +137,13 @@ impl Tensor {
         if self.is_contiguous() {
             return self.clone();
         }
-        // Copy to contiguous layout on CPU
-        assert_eq!(self.device(), Device::Cpu, "contiguous() on GPU not yet supported");
+        // For GPU tensors: round-trip through CPU (correct but slow).
+        // TODO: write a GPU contiguous-copy kernel for performance.
+        if matches!(self.device(), Device::Cuda(_)) {
+            let cpu = self.to_device(Device::Cpu);
+            let contig = cpu.contiguous();
+            return contig.to_device(self.device());
+        }
         let numel = self.numel();
         let elem_size = self.dtype.size_bytes();
         let src_bytes = self.storage.as_cpu_bytes();
@@ -173,17 +178,18 @@ impl Tensor {
     // --- Device transfer ---
 
     pub fn to_device(&self, device: Device) -> Self {
-        let t = if self.is_contiguous() { self.clone() } else { self.contiguous() };
-        if t.device() == device {
-            return t;
+        if self.device() == device {
+            return self.clone();
         }
-        let new_storage = t.storage.to_device(device).expect("device transfer failed");
+        // Transfer the raw storage (preserving strides/offset).
+        // Non-contiguous layout is preserved — the user can call contiguous() after.
+        let new_storage = self.storage.to_device(device).expect("device transfer failed");
         Self {
             storage: new_storage,
-            shape: t.shape,
-            strides: t.strides,
-            offset: 0,
-            dtype: t.dtype,
+            shape: self.shape.clone(),
+            strides: self.strides.clone(),
+            offset: self.offset,
+            dtype: self.dtype,
         }
     }