Over in #6573 I added support for optional dtypes via specifying an optional<ScalarType>=nullopt parameter, but this fails in the jit because the jit doesn't know how to handle ScalarTypes and/or optional values.

For now I believe I can break these up into two separate functions, with a couple of downsides:

1. Can't trace the dtype version
2. The python arg parser will generate two different overloads (which leads to slower parsing and ugly error messages) even though there is really only a single function.

CC @zdevito @apaszke.