Skip non-numeric covariates when computing standardized differences

[arpitjain099]

_get_numeric_vars is documented as working out which covariates are numeric, but it never looks at the dtype. A string column with at least threshold distinct values passes the continuous test, and the standardized difference then dies inside numpy:

X = pd.DataFrame({'age': ..., 'group': rng.choice(list('abcdef'), size=n), 'flag': ...})
_get_numeric_vars(X)     # (['age', 'group'], ['flag'])
get_std_diffs(X, w)      # TypeError: ufunc 'divide' not supported for the input types

That reaches users through plot_ps_diagnostics, which takes whatever covariate columns the caller names, and categorical covariates are a normal thing to have in a propensity model. The function already has the right behaviour for columns it cannot use: it drops them and logs which ones. Non-numeric columns now take that path, so the diagnostic reports balance for the covariates it can measure instead of failing outright.

The dropped-variables message was only describing two of its three reasons, so it now mentions the non-numeric case too.

Two tests: a frame with a string covariate classifies as ['age'] continuous and ['flag'] binary and produces finite differences, and a frame whose only covariate is non-numeric still raises the existing ValueError for having nothing to measure. The first fails on master with assert ['age', 'group'] == ['age'] and then the TypeError above.

pytest tests/test_visualize.py tests/test_visualize_score_ci.py is 33 passing, and black is clean.