Fix alloc-dealloc-mismatch in <str_vec> destructor; drop unused DELTA token

[Licht-T]

Fixes the heap corruption observed as envoyproxy/envoy#36471 (postgres_proxy crash under pgbench).

Bugs

1. <str_vec> %destructor uses delete on malloc-ed pointers

The shared destructor at bison_parser.y#L147-L154 covers <str_vec> together with <table_vec>, <column_vec>, etc., and calls delete ptr on each element. <str_vec> elements are char* produced by the lexer via strdup() (unquoted IDENTIFIER) or hsql::substr() (quoted IDENTIFIER, also malloc-backed). Mixing free/delete is undefined behavior. Under tcmalloc with -fsized-deallocation, sized operator delete(void*, size_t) trusts the static type size (1 for char) and returns the chunk to the wrong size-class freelist; after enough mismatched frees an unrelated allocation segfaults.

This is the same bug already fixed upstream in hyrise/sql-parser#221 (Oct 2022). This fork was branched before that PR landed and never resynced.

AddressSanitizer flags it cleanly:

ERROR: AddressSanitizer: alloc-dealloc-mismatch (malloc vs operator delete)
    #0 operator delete(void*)
    #1 yydestruct                bison_parser.y:156
    #2 hsql_parse
    ...
allocated by:
    #0 __interceptor_strdup
    #1 hsql_lex                  flex_lexer.l:250

2. DELTA is a dead reserved keyword

%token DELTA is declared (and listed in sql_keywords.txt#L95 and flex_lexer.l#L128) but referenced by no grammar rule. The lexer therefore tokenizes delta as DELTA and the parser has nowhere to consume it, so any SQL using delta as an identifier (notably pgbench's pgbench_history.delta) fails to parse - which is precisely what triggers the destructor cleanup path that exhibits bug #1.

Fix

• Split the <str_vec> destructor out of the shared block and use free() for its elements. Keep delete for the other vector types (which hold new'd objects).
• Drop DELTA from sql_keywords.txt, bison_parser.y (%token declaration), and flex_lexer.l (lexer rule).
• Regenerate the committed bison_parser.{cpp,h} and flex_lexer.{cpp,h} accordingly. (This accounts for the bulk of the diff — the hand-written changes are small.)

Tests

test/regression_tests.cpp adds:

• DeltaIsAValidIdentifier - INSERT INTO pgbench_history (tid, bid, aid, delta) VALUES (1, 2, 3, 4) parses to a valid InsertStatement with delta as the fourth column.
• RepeatedFailedInsertParseDoesNotCorruptHeap - runs an INSERT whose column list contains a still-reserved keyword 1000× to exercise the destructor cleanup path. Under ASAN this catches the regression on the very first iteration; in release builds it serves as a smoke test.

Verified:

• make test: all 87 grammar tests + 12 unit tests + 2 new regression tests pass.
• valgrind memory-leak check clean.
• Grammar conflict check clean.
• ASAN build clean against the new tests; reverting the destructor reproduces the exact diagnostic above.
• Built envoyproxy/envoy main against this branch via --override_repository and reran the issue's pgbench scenario (50 clients × 30 s, plaintext, enable_sql_parsing: true) - no crash.