The first thing means that
$$ \partial_\bar{z} \langle T(z) O_1(z_1,\bar{z_1}) O_2(z_2,\bar{z_2}) ... O_n(z_n,\bar{z_n}) \rangle = 0$$
except when $z=z_i$ for some i (coinciding insertion point for T and some other operator), and $O_i$ is any local operator at a given point (not necessarily holomorphic).
This Ward identity is an operator equation, it is like an equation of motion, it tells you that the stand-alone operator T(z) has a zero z-bar derivative. This means it has a zero z-bar derivative in any state, and the operators $O_i$ just prepare the state, and whether they are holomorphic or anti-holomorphic, the z-bar derivative of T is zero, so the z-bar derivative of the correlation function is zero.
When there is a coincidence of positions, you have ordering issues preventing the naive equation of motion from holding, because the path integral time-orders all operators automatically. So the derivative includes delta-function terms which reorder the thing, and the equation fails. The holomorphic Ward identity is stronger than this in its actual statement: it says that the right hand side determines the holomorphic transformation properties of each of the local fields.
This implies that if you expand T(z)O(z') in an OPE, whether O is holomorphic or not, the coefficients must be holomorphic away from coinciding points, since
$$T(z) O(z') = \sum_{k=-n}^{\infty} C_k(z-z') O_k(z) $$
And the zero value of the z-bar derivative of the left-hand side means that the z-bar derivative of the right hand side, and therefore of the coefficient functions, is also zero except when z=z'.