Initial project scaffold

tasks/02_row_softmax/cuda_skeleton.cu

@@ -0,0 +1,11 @@
+// Workbook-local CUDA sketch for row softmax.
+//
+// Reflection prompt:
+// Softmax is usually bandwidth-bound because the math is cheap but the rows are read and written a lot.
+// Keep track of how many global-memory passes your implementation needs.
+
+// TODO(student):
+// 1. Assign one block or block tile to a row.
+// 2. Compute the row max.
+// 3. Compute the sum of exp(x - row_max).
+// 4. Normalize the row.