Fix bug due to precision in CategoricalMatrix._get_col_stds

[mlondschien]

This fixes a bug in CategoricalMatrix._get_col_stds: We observed that if using float32 precision, we'd sometimes get nans as standard deviations for constant categorical columns. This would then lead to nan in P1 and P2 here.

To reproduce:

import numpy as np
import tabmat
import pandas as pd
import glum

n = 1_234_456
rng = np.random.default_rng(0)
p = 0.00001

x = rng.choice([0, 1], p=[p, 1-p], size=n)
cat = pd.Categorical.from_codes(x, categories=np.array(range(3), dtype=np.int32))
mat = tabmat.CategoricalMatrix(cat, column_name="cat", dtype=np.float32)

y = np.random.choice([0, 1], size=n).astype(np.float32)

weights = np.ones(n, dtype=np.float32)
weights /= weights.sum()

X, means, col_stds = mat.standardize(weights, True, True)
print(f"means: {means}, col_stds: {col_stds}")

glm = glum.GeneralizedLinearRegressor(family="binomial", alpha=0.01, l1_ratio=0.5)
glm.fit(mat, y)

/Users/mlondschien/mambaforge/envs/icufm/lib/python3.10/site-packages/tabmat/categorical_matrix.py:607: RuntimeWarning: invalid value encountered in sqrt
  return np.sqrt(mean - col_means**2)
means: [9.7208804e-06 1.0003393e+00 0.0000000e+00], col_stds: [0.00311782        nan 0.        ]
/Users/mlondschien/mambaforge/envs/icufm/lib/python3.10/site-packages/tabmat/categorical_matrix.py:607: RuntimeWarning: invalid value encountered in sqrt
  return np.sqrt(mean - col_means**2)
Traceback (most recent call last):
  File "/Users/mlondschien/code/projects2024-icu-foundation/tmp_tabmat.py", line 22, in <module>
    glm.fit(mat, y)
  File "/Users/mlondschien/mambaforge/envs/icufm/lib/python3.10/site-packages/glum/_glm.py", line 3244, in fit
    coef = self._solve(
  File "/Users/mlondschien/mambaforge/envs/icufm/lib/python3.10/site-packages/glum/_glm.py", line 1113, in _solve
    coef, self.n_iter_, self._n_cycles, self.diagnostics_ = _irls_solver(
  File "/Users/mlondschien/mambaforge/envs/icufm/lib/python3.10/site-packages/glum/_solvers.py", line 288, in _irls_solver
    state.eta, state.mu, state.obj_val, coef_P2 = _update_predictions(
  File "/Users/mlondschien/mambaforge/envs/icufm/lib/python3.10/site-packages/glum/_solvers.py", line 638, in _update_predictions
    return eta_mu_objective(
  File "/Users/mlondschien/mambaforge/envs/icufm/lib/python3.10/site-packages/glum/_solvers.py", line 671, in eta_mu_objective
    obj_val += linalg.norm(P1 * coef[intercept_offset:], ord=1)
  File "/Users/mlondschien/mambaforge/envs/icufm/lib/python3.10/site-packages/scipy/linalg/_misc.py", line 146, in norm
    a = np.asarray_chkfinite(a)
  File "/Users/mlondschien/mambaforge/envs/icufm/lib/python3.10/site-packages/numpy/lib/function_base.py", line 628, in asarray_chkfinite
    raise ValueError(
ValueError: array must not contain infs or NaNs

Not sure where to test.