BUG: `LogisticRegression.fit` has poor performance on speed

[JiaYaobo]

Describe the bug

LogisticRegression.fit never stops with a bit larger data.

To Reproduce

When max_iter=1 everything works fine

from xorbits._mars.learn.glm import LogisticRegression
import numpy as np

n_rows = 100
n_cols = 2
X = np.random.randn(n_rows, n_cols)
y = np.random.randint(0, 2, n_rows)

lr = LogisticRegression(max_iter=1)
lr.fit(X, y)

However, just increase max_iter to 100, the program seems never stop (at least after 1min, it's weird.)

lr = LogisticRegression(max_iter=100)
lr.fit(X, y)

1. Your Python version: 3.10.2
2. The version of Xorbits you use: HEAD, install on my local device.
3. I'm working on my Macbook with m1 pro chip

[JiaYaobo]

After some inspection, https://github.com/mars-project/mars/issues/2505#issue-1021662005 perhaps explain it, gradient based loop solution for logistic regression is inefficient for xorbits.

[luweizheng]

Maybe we should use a different approach on the control flows, like for loop. For example, the method in this paper.