32-bit floating point warnings. (#278)

* 32-bit floating point warnings.

* Add conditioning causes comment to readme.

README.md

@@ -130,6 +130,9 @@ print('Model log-likelihood', get_obj_val(dat, 'poisson', 0.0, 0.0, model.interc
 
 [This is a brief tutorial on Tweedie Regression with L2 regularization from sklearn. `quantcore.glm` has many more features and capabilities but it can also replicate everything done here.](https://scikit-learn.org/stable/modules/linear_model.html#generalized-linear-regression)
 
+## A quick note on 32-bit floating point precision
+Note that a float32 data matrix is acceptable and will result in the entire algorithm being run in 32-bit precision.  However, for problems that are poorly conditioned, this might result in poor convergence or flawed parameter estimates. Poor conditioning can be caused by a large number of features with few observations, highly collinear features and by light regularization. In spite of this risk, it's common to see 1.5-2x improvements in runtime with 32-bit data.
+
 # Testing/Continuous integration
 
 ## Golden master tests

src/quantcore/glm/_glm.py

@@ -1685,7 +1685,10 @@ def fit(
         Parameters
         ----------
         X : {array-like, sparse matrix}, shape (n_samples, n_features)
-            Training data.
+            Training data. Note that a float32 matrix is acceptable and will
+            result in the entire algorithm being run in 32-bit precision.
+            However, for problems that are poorly conditioned, this might result
+            in poor convergence or flawed parameter estimates.
 
         y : array-like, shape (n_samples,)
             Target values.