Test improvements and minor code fixes for formatting.c.

Over in the thread about teaching to_number() to handle Roman
numerals [1], it was noted that we currently have precisely zero
test coverage for to_char()'s existing Roman-numeral code, except
in the timestamp code path which shares nothing with the numeric
code path.

In looking at this, I found that there's also no test coverage
for the EEEE, V, or PL format codes. Also, the possibility of
overflow while converting an input value to int in order to
pass it to int_to_roman was ignored. Attached is a patch that
adds more test coverage and cleans up the Roman-numeral code
a little bit.

BTW, I also discovered that there is a little bit of support
for a "B" format code: we can parse it, but then we ignore it.
And it's not documented. Oracle defines this code as a flag
that:

Returns blanks for the integer part of a fixed-point number
when the integer part is zero (regardless of zeros in the
format model).

It doesn't seem super hard to implement that, but given the
complete lack of complaints about it being missing, maybe we
should just rip out the incomplete support instead?

regards, tom lane

[1] https://www.postgresql.org/message-id/flat/CAMWA6ybh4M1VQqpmnu2tfSwO%2B3gAPeA8YKnMHVADeB%3DXDEvT_A%40mail.gmail.com