drivers: serial: ns16550: fill full fifo

Put the maximum number of bytes into Tx FIFO in the fill_fifo routine
to reduce CPU usage.

Previously, the THRE bit was checked in a loop, but, according to doc -
"In the FIFO mode, it is set when the XMIT FIFO is empty, and is
cleared when at least one byte is written to XMIT FIFO.", so only one
byte was transferred every interrupt. That was generating a big amount
of interrupts, which consumes CPU time.

Signed-off-by: Dawid Niedzwiecki <dn@semihalf.com>