mps: clone() aliases its source, so x.clone().add_(1) changes x #97

Measured on this Mac with the real shim (develop and feat/mps-index alike): on mps, y = x.clone(); y.add_(1) changes x too; so do fill_, setitem, clone(memory_format=contiguous) and detach().clone(). cpu is correct. No compute or blit encoder runs: nothing is copied.

Cause: clone_default calls candle Tensor::copy(), whose MetalStorage::try_clone is upstream's Ok(self.clone()), an Arc<Buffer> share. candle never writes a buffer in place, so it was safe there; torch._C does.

Done when

• MetalStorage::try_clone in the candle fork copies into a fresh buffer (one blit), patch regenerated, vendor_candle.sh --check passes.
• A test: every clone spelling x every in-place write leaves the source unchanged on mps; detach and alias still share (upstream semantics); a blit encoder is counted. Red before.