saner collect_arg() code generation

1) gently coax gcc code generator away from screwing the bitfields
handling it is so keen to do - pass struct position from the caller
by value, with .newline already cleared (with inlined helper that
does trimming), do *not* let collect_arg() see that it's always
setting ->pos.newline to 0 (otherwise the FPOS tries to peel the damn
thing apart, turning mask-then-or into a bleeding horror).

2) get rid of 'vararg' checks for commas by changing the way we
keep track of nesting parentheses - instead of
	number of '(' - number of ')'
use
	2 * number of '(' - 2 * number of ')' + vararg
That still recognizes the closing parenthesis of argument list
(counter gets negative there, vararg or no vararg) and it does
recognize an argument-terminating comma without having to look
at 'vararg' - just check if counter is zero.
Reduces spills nicely.

3) do *not* use eof_token() as loop termination condition, compare
token_type() with TOKEN_EOF instead.  Yes, it's faster to compare
the pointer with constant instead, but it is *not* going to match
that unless we have an unterminated argument list and we are going
to check for TOKEN_STREAM{BEGIN,END} immediately afterwards anyway,
with the same "leave the loop now" on match, so these checks combine
nicely.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>