Handle NaN values in Y normalization

[micchu]

Since the results of Y normalization can contain not only np.inf but also np.nan, I added a process to replace them with 0.

Numpy operations have the following properties:

• 1/0 results in np.inf (dividing a non-zero value by 0)
• 0/0 results in np.nan (dividing 0 by 0)

Since this occurs immediately after Y - y_mean, the procedure should essentially be 0/0. However, a very small error in float32 results in an operation like 3.2e-25/0 instead of 0/0, resulting np.inf in most cases. This makes this problem difficult to notice.
In very rare cases, 0/0 occurs, resulting in np.nan feature values. In such cases, the weights W contain np.nan values. This causes np.nan values ​​to appear in the predicted values, causing problems when using the predicted values ​​for reconstruction, etc.

[KenyaOtsuka]

I do not think this should be accepted as-is.

The proposed change treats every NaN after Y normalization as if it came from 0/0, but that is not necessarily true. NaN may also indicate invalid input data or invalid normalization parameters. In those cases, silently replacing them with 0 may hide invalid input data or normalization parameters.

A safer fix would be to handle only the y_norm == 0 case explicitly, and then check that the normalized Y contains no remaining non-finite values.