It is a mistake for the lexer not to make a copy of regexp literals,
because the parser may need to shift more than one of them onto its
stack as it searches for a matching rule.

For the following demonstration, the correct output is "1b":

        $ echo ab | ./master 'sub(/a/, "b" ~ /b/)'
        a1

The debugging output confirms that /b/ clobbered /a/, incorrectly
modifying sub's first argument:

        $ echo ab | ./master -d 'sub(/a/, "b" ~ /b/)' | grep reparse
        reparse <b>
        reparse <b>

One of those should have been "reparse <a>".

I introduced this bug in a fix [1] for memory leaks discovered by
Todd C. Miller [2]. This commit reverts my fix and applies a slightly
modified version of Todd's recommendation (I omit the INDEX case
because reg_expr is stored in a node).

[1] 577a67cc
[2] https://github.com/onetrueawk/awk/pull/156

Testing with `valgrind --leak-check=full $awk "$prog" </dev/null`, the
leaks produce something like this:

==473055== 4 bytes in 1 blocks are definitely lost in loss record 36 of 128
==473055==    at 0x484586F: malloc (vg_replace_malloc.c:381)
==473055==    by 0x49E54FD: strdup (strdup.c:42)
==473055==    by 0x409288: tostring (tran.c:522)
==473055==    by 0x411DF5: regexpr (lex.c:557)
==473055==    by 0x402E99: yyparse (awkgram.tab.c:2251)
==473055==    by 0x402877: main (main.c:211)

Where $prog is any one of these one-liners:

'/abc/; /cde/'
'"abc" ~ /cde/; /fgh/'
'match(/abc/, /cde/)'
'split(/abc/, a, /cde/)'
'sub(/abc/, /cde/)'
'gsub(/abc/, /cde/)'
'sub(/abc/, /cde/, a)'
'gsub(/abc/, /cde/, a)'