runtest first tries *without* the given extra-cflags, so -MMD was never added. singletest, instead, always use the args

it's internal documentation only, not meant to be distributed.

if some of the tp_* function fails, the inner code could free the template too (that's what galileo does), so `tp->tp_ret' would access free'd memory! Instead, add a local variable `tp_ret' to keep track of the return code. While here, drop the `render' shortand that adds the argument call, the template_reset function and adapt the regress.