Implement all 5 Triton kernel labs

- vector_add: basic masked load/store with block indexing
- row_softmax: single-pass numerically stable softmax per row
- tiled_matmul: K-dimension tile loop with edge masking (IEEE precision)
- online_softmax: two-pass running max/sum recurrence across blocks
- flash_attention_fwd: blockwise Q/K/V with online softmax, causal support

All 26 tests pass on RTX 5090 (CUDA 12.8, Triton 3.6).

kernels/triton/vector_add.py

@@ -26,10 +26,10 @@ if TRITON_AVAILABLE:
         pid = tl.program_id(axis=0)
         offsets = pid * block_size + tl.arange(0, block_size)
         mask = offsets < num_elements
-        # TODO(student): load x and y using masked tl.load calls.
-        # TODO(student): add the vectors.
-        # TODO(student): write the result with tl.store.
-        pass
+        x = tl.load(x_ptr + offsets, mask=mask)
+        y = tl.load(y_ptr + offsets, mask=mask)
+        out = x + y
+        tl.store(out_ptr + offsets, out, mask=mask)
 
 
 def triton_vector_add(x: torch.Tensor, y: torch.Tensor, block_size: int = 1024) -> torch.Tensor:
@@ -40,5 +40,9 @@ def triton_vector_add(x: torch.Tensor, y: torch.Tensor, block_size: int = 1024)
         raise ValueError(f"shape mismatch: {x.shape} vs {y.shape}")
     if not x.is_cuda or not y.is_cuda:
         raise ValueError("Triton kernels in this lab expect CUDA tensors.")
-    raise NotImplementedError("TODO(student): launch vector_add_kernel and return the output tensor.")
+    out = torch.empty_like(x)
+    num_elements = x.numel()
+    grid = ((num_elements + block_size - 1) // block_size,)
+    vector_add_kernel[grid](x, y, out, num_elements, block_size=block_size)
+    return out