ENH: np.append should cast appended values to the dtype of the first argument

[ngoldbaum]

Proposed new feature or change:

Right now np.append will convert both operands to an array first, then combine the result. This produces suboptimal results for stringdtype and a scalar operand:

Pdb) np.append(stringdtype_array, "hello")
*** ValueError: Cannot find common instance for unequal dtype instances

If instead we cast the second argument to the dtype of the first argument, this would succeed.

Note that this problem is particular to StringDType, but relying on numpy coercing the arguments to built-in dtypes is probably less forward looking than casting to the dtype of the first argument.

[seberg]

See also gh-22429. While I agree that append shouldn't promote, I am not certain how well we can get away with it.

I am surprised about not allowing the promotion here, but I guess it is the conservative thing to do (easier to relax).