Why can't we feed datetime to Linear Regression and how does toordinal() different from any other integer datatype?

Linear Regression is associating any numerical (or binary, which is a particular numerical) value to a coefficient. Multiplying those values by those coefficients gives you an output, and setting the threshold, you know if the model predicts 1 or 0. (This is a brief summary, you'll find plenty of people explaining in details how it works).

If your variable is a date, you have a format, like "Year/Month/Day", and the regression doesn't know how to interpret it, since it need numerical data.

So as to use the date in a regression, you can basically take the year, as a variable. If you want to be more precise, you'll have to create your own variable, with the meaning you want (Ex : a binary variable taking 1 if the day is a Saturday or Sunday, 0 either).

Side Note : Watch out when you use dates variables, if you want to use your model to predict new, future data. For ex, if you have a variable Year, and you test your model in a Cross-Validation, it will have an impact since the trend can change with years. Moreover, if you want to predict new values input, Year will always be '2020', so this won't be that useful. Utilizing it to evaluate your model might biaise your results.