The mantissa in USART_BRR is 12-bit i.e. its maximum value is 4095. That limits accordingly the minimum attainable baudrate for given APB frequency.
The following values in example tables are thus invalid and those baudrates (1200/2400Baud, quite common in various environments) can't be achieved at those particular APB-clock setups.
Found out the hard way - on a 16-sampling USART1 at the maximum APB2 the USART supposed to work at 1200 Baud was working at almost 19200 Baud, as the mantisa of baudrate divider was 0x117 instead of the required 0x1117.
Instead of those lengthy and quite useless tables, those limits should be highlighted in the UM, and formulas to calculate the limits and the errors should be given; and then a more detailed explanation should go to a well-crafted AN, which can then house the extensive tables, too, and even more. My 2 eurocents.