The non-CUDA TapeMachine exec switch checked usePrealloc (a runtime
check for non-nil destination register) before instr.useUnsafe (a
compile-time flag for shared input/output registers). When registers
are shared, the destination always has a value (the input), so
usePrealloc was always true, shadowing the useUnsafe case entirely.

For ops that implement UnsafeDo but not UsePreallocDo (e.g.
elemUnaryOp), this caused plain Do() to run instead of UnsafeDo(),
allocating new memory rather than modifying in-place. This defeated
the register sharing optimization.

The CUDA execution path (vm_tape_cuda.go) has no usePrealloc case
and already checks useUnsafe directly after preAllocated. This change
aligns the non-CUDA path with the CUDA path's intended design.