Extend our type signatures of inner and dot to match NumPy's types.

[adamreichold]

Closes #284

[kngwyu]

Oops, I didn't notice that this can cause an error, thanks.
Could you please some test cases that extract floats/integers also in (non-doc) tests?

Also, can we do the same for einsum?

[adamreichold]

Oops, I didn't notice that this can cause an error, thanks.

Yeah, I think the effective signature of the capsule API is quite unfortunate considering a zero-dimensional array can be used almost exactly like a scalar in Python code.

Could you please some test cases that extract floats/integers also in (non-doc) tests?

Added more tests for inner and dot as a separate commit.

Also, can we do the same for einsum?

Indeed this implies here too depending on the summation performed. Added as a separate commit.

[adamreichold]

As this break the API anyway, I took the liberty of renaming the einsum_impl function to einsum. There is no danger of conflicts as macros and functions inhabit distinct namespaces and most people will invoke the macro anyway. But I think the _impl suffix makes the name look unnecessary "internal."

[adamreichold]

Slipped in a further change to make use of PyArray::item to basically game our coverage up... |-\

[adamreichold]

@kngwyu Are you satisfied with the adjustments?

[kngwyu]

Sorry for being late. Thank you for addressing my requests!