further fix for RT #23810
The fix for #23810,
27e904532594b7fb, introduced a regression, spotted by
Nicholas as RT #75146.
Basically, in S_doeval() if the yyparse() fails due to dying (rather than
just bailing out with with a syntax error, say), then the topmost EVAL
context will have been popped. My improved error handling code mostly
understood the difference, but forgot that in the die case, PL_eval_root
will have been restored to its previous value by the CX pop, and thus
its value shouldn't be messed with.