phase 2: tensor abstraction layer

- DType enum (F32, F16, BF16) with TensorDType trait
- Shape utilities: contiguous_strides, broadcast_shape, broadcast_strides
- Storage with Arc reference counting (CPU Vec<u8> or GPU GpuBuffer)
- Device enum (Cpu, Cuda(id)) with to_device transfer
- Tensor type with strided layout: reshape, transpose, squeeze, unsqueeze
- contiguous() copies non-contiguous views to contiguous layout
- from_slice, zeros, ones constructors
- as_slice<T> for typed CPU read access, data_ptr for GPU kernel launch
- CPU↔GPU roundtrip verified
- All 27 tests pass (12 cuda + 4 shape + 11 tensor)

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

crates/xserv-tensor/src/storage.rs

@@ -0,0 +1,119 @@
+use std::sync::Arc;
+use xserv_cuda::{GpuBuffer, Result as CudaResult};
+
+enum StorageInner {
+    Cpu { data: Vec<u8> },
+    Cuda { buffer: GpuBuffer },
+}
+
+/// Reference-counted storage for tensor data. Multiple tensors can share
+/// the same storage (e.g., after transpose or slice — view semantics).
+#[derive(Clone)]
+pub struct Storage(Arc<StorageInner>);
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Device {
+    Cpu,
+    Cuda(u32),
+}
+
+impl std::fmt::Display for Device {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Device::Cpu => write!(f, "cpu"),
+            Device::Cuda(i) => write!(f, "cuda:{i}"),
+        }
+    }
+}
+
+impl Storage {
+    pub fn cpu(data: Vec<u8>) -> Self {
+        Self(Arc::new(StorageInner::Cpu { data }))
+    }
+
+    pub fn cuda(buffer: GpuBuffer) -> Self {
+        Self(Arc::new(StorageInner::Cuda { buffer }))
+    }
+
+    pub fn device(&self) -> Device {
+        match self.0.as_ref() {
+            StorageInner::Cpu { .. } => Device::Cpu,
+            StorageInner::Cuda { .. } => Device::Cuda(0),
+        }
+    }
+
+    pub fn len_bytes(&self) -> usize {
+        match self.0.as_ref() {
+            StorageInner::Cpu { data } => data.len(),
+            StorageInner::Cuda { buffer } => buffer.len(),
+        }
+    }
+
+    /// Get a read-only view of CPU data. Panics if storage is on GPU.
+    pub fn as_cpu_bytes(&self) -> &[u8] {
+        match self.0.as_ref() {
+            StorageInner::Cpu { data } => data,
+            StorageInner::Cuda { .. } => panic!("cannot access GPU storage as CPU bytes"),
+        }
+    }
+
+    pub fn gpu_buffer(&self) -> &GpuBuffer {
+        match self.0.as_ref() {
+            StorageInner::Cuda { buffer } => buffer,
+            StorageInner::Cpu { .. } => panic!("cannot access CPU storage as GPU buffer"),
+        }
+    }
+
+    /// Copy to a different device. If already on the target device, clones the Arc (no copy).
+    pub fn to_device(&self, target: Device) -> CudaResult<Self> {
+        let current = self.device();
+        if current == target {
+            return Ok(self.clone());
+        }
+        match (current, target) {
+            (Device::Cpu, Device::Cuda(_dev)) => {
+                let cpu_data = self.as_cpu_bytes();
+                let mut buf = GpuBuffer::alloc(cpu_data.len())?;
+                buf.copy_from_host(cpu_data)?;
+                Ok(Storage::cuda(buf))
+            }
+            (Device::Cuda(_), Device::Cpu) => {
+                let gpu_buf = self.gpu_buffer();
+                let mut data = vec![0u8; gpu_buf.len()];
+                gpu_buf.copy_to_host(&mut data)?;
+                Ok(Storage::cpu(data))
+            }
+            (Device::Cuda(_), Device::Cuda(_)) => {
+                let src = self.gpu_buffer();
+                let mut dst = GpuBuffer::alloc(src.len())?;
+                dst.copy_from_device(src)?;
+                Ok(Storage::cuda(dst))
+            }
+            _ => unreachable!(),
+        }
+    }
+
+    /// Create a new owned copy of the storage on the same device.
+    pub fn deep_copy(&self) -> CudaResult<Self> {
+        match self.0.as_ref() {
+            StorageInner::Cpu { data } => Ok(Storage::cpu(data.clone())),
+            StorageInner::Cuda { buffer } => {
+                let mut dst = GpuBuffer::alloc(buffer.len())?;
+                dst.copy_from_device(buffer)?;
+                Ok(Storage::cuda(dst))
+            }
+        }
+    }
+
+    /// Allocate zeroed storage on the given device.
+    pub fn zeros(len_bytes: usize, device: Device) -> CudaResult<Self> {
+        match device {
+            Device::Cpu => Ok(Storage::cpu(vec![0u8; len_bytes])),
+            Device::Cuda(_) => {
+                let mut buf = GpuBuffer::alloc(len_bytes)?;
+                buf.zero()?;
+                Ok(Storage::cuda(buf))
+            }
+        }
+    }
+}