fix: keep torch_scalar_tensor() a leaf when requires_grad = TRUE

[taekop]

torch_scalar_tensor() created the tensor with requires_grad = TRUE and then called $squeeze() on it. Since $squeeze() is a tracked autograd op, the returned tensor was no longer a leaf, so .grad was never populated after backward() (as sebffischer diagnosed in the issue thread).

This fix squeezes first and only enables requires_grad afterwards via $requires_grad_(), for the non-tensor-input branch (the tensor-input branch already called $requires_grad_() last and was unaffected). dtype/device handling is unchanged.

Added a regression test covering both branches (plain value and tensor input) that checks is_leaf and the gradient value after backward().

Fixes #1368.