Support `hl.arange()` with non-power-of-2 input

[yf225]

When hl.arange() is used with non-power-of-2 input, it currently throws error. But ideally we should be able to support it by applying automatic padding (similar to how we do tensor factory padding for non-power-of-2 shapes #804).

(@mengluy0125 first raised this feature request in internal forum.)