RFC: avoid deprecation warnings from numpy 2.5 when updating an array's dtype in place

[neutrinoceros]

close yt-project/yt#5420

[ngoldbaum]

Did you try deleting the code that updates the dtype and allow integer arrays through unchanged? Maybe it just works?

Although it occurs to me now that maybe this code is setting up an in-place multiply that needs a float array, since multiplying by a non-integral conversion factor will suffer truncation error if the array stays an integer array.

Another possibly more disruptive option is to raise a TypeError so callers can reason about this.

[neutrinoceros]

Did you try deleting the code that updates the dtype and allow integer arrays through unchanged? Maybe it just works?

It doesn't. 15 tests break quite badly with errors like

numpy._core._exceptions._UFuncOutputCastingError: Cannot cast ufunc 'divide' output from dtype('float64') to dtype('int64') with casting rule 'same_kind'

Another possibly more disruptive option is to raise a TypeError so callers can reason about this.

I agree that's probably where we should be going with this, though we could probably do that over a deprecation cycle.