Skip missing values when center() and standardize() memorize their statistics

[raashish1601]

Fixes #208, and the missing-value case of #166.

center() and standardize() memorize their mean (and variance) over every row, including missing ones. A single NaN in the column makes the mean NaN, so the transformed column is NaN in every row, and the NA action then drops all rows:

df = pd.DataFrame({"x1": np.arange(5), "x2": [1, 2, np.nan, 5, 6]})
dmatrix("~ x1 + standardize(x2)", df)
# DesignMatrix with shape (0, 3)

With this change the statistics skip missing values column by column, as R's scale() does (colMeans(x, na.rm = TRUE)), and the transform leaves NaN where the input was missing. The NA action then drops only the rows that actually have missing values:

[[ 1.          0.         -1.21267813]
 [ 1.          1.         -0.72760688]
 [ 1.          3.          0.72760688]
 [ 1.          4.          1.21267813]]

Inputs without missing values give the same results as before. Only float/complex inputs are checked for NaN; other dtypes keep the previous path.

Tests: added test_Center_missing_values and test_Standardize_missing_values to patsy/test_state.py, with two columns that have NaN in different rows, memorized in one chunk and in two chunks. They fail on main and pass with this change. The full test suite passes locally and ruff format is clean.

[bashtage]

patsy is in effectively long-term maintenance mode with mostly only seeing absolutely required fixes to stop it from failing to operate as it has been on new versions of python.

I wouldn't say this is a hard no, but just to set expectations that we mostly don't take features or bug fixes

[raashish1601]

Understood, thanks for setting expectations. I'll leave it open in case it's useful, and I'm happy to close it if you'd rather keep patsy as is.