Don't call setjmp in code we didn't write.

Don't call setjmp in code we didn't write.

Using setjmp() in a routine requires that anything whose value needs
*not* to be restored to its value when setjmp() was called in a
longjmp() be declare "volatile".

We can't force Bison or Berkeley YACC to do that with variables in the
parser function, so we can't safely do a setjmp() in the parser
function.  *Some* compilers might recognize setjmp() and automatically
do that, either silently or with a warning, but that's not guaranteed by
the C language specification.

This could cause a problem if it trashes the value of local variables
storing pointers to the parser's pushdown stack, if they're assumed to
point to the *current* stack at the time the stack is freed at the end
of the parser function.

Instead, use setjmp/longjmp only inside functions defined in gencode.c;
have all functions called by the parser do a setjmp and, if it returns
1, return a null pointer, and have all those calls check the return
value and, if it's null, do a YYABORT.

Add a bpf_set_error() routine, for use *outside* gencode.c, which just
sets the error string.  In the parser, do a YYABORT after calling it;
in the lexical analyzer, return a token even for errors, but make sure
the token will cause the parse to stop.

Credit to OSS-Fuzz for possibly finding this issue (it may be what's
causing crashes in some tests).