Changelog

A release about writes that went nowhere, and the CPU kernels that were slower than they had to be. The engine is 0.16.1 with ABI 16 unchanged, so a C++ extension built against 3.15.4 keeps working. - On Metal, an in-place write through a view reached nothing but the view. x[0].add_(1), x[:, 0].zero_(), x.t()[1].mul_(2), r[1][0] = v, p.data.sub_(lr * g), x.detach().add_(...) and p.grad.mul_(s) all left the tensor they came from unchanged, with no error, while the CPU and the reference wrote through. A hand-written optimizer step on Metal trained nothing. Writes now go back to the base, a view of a view all the way up, with gradients identical to the CPU's. Overlapping views and views of a leaf that requires grad are refused, as on the CPU. A view taken *before* its base is written still does not see that write on Metal. - `with no_grad(): weight[i] = v` after a forward lost the weight's gradient on both devices. The assignment swapped a new tensor under the Parameter while the graph kept the old one. It now writes in place, so backward reports that the saved weight changed, as the reference does. - CPU `gather` and `index_select` are 8 to 14 times faster, and a bicubic resize is 4 times faster (375×500 → 600×800: 132 → 33 ms on an M1 Pro). - The tanh-approximate GELU is 34 times faster on the CPU (4.1 → 0.12 ms on 524k elements). It had been ten times slower than the exact form it approximates. The exact GELU is 34% faster: float32 erf is now a vectorised polynomial within one ulp of the correctly rounded value, and parallel CPU loops now balance across performance and efficiency cores instead of waiting on the slowest. Behaviour that changes: on Metal, a write through a view now changes the base, as it always did on the CPU; weight[i] = v under no_grad after a forward makes the next backward raise instead of silently dropping the weight's gradient; a Core ML export accepts an input that requires grad and is written by the forward; and float32 erf on the CPU can differ from the previous build by one ulp.