fix: CREATE TABLE does not emit sqlite_autoindex_* for declared composite keys (oracle calls the file malformed)

[dpsiderius]

Description

The other half of #685. That ticket fixed adopting a stock-created
file: an existing sqlite_autoindex_* is now recovered and maintained.
This one fixes producing one.

CREATE TABLE with a declared composite PRIMARY KEY or UNIQUE
constraint currently creates no sqlite_autoindex_* b-tree and no
sqlite_master row for it. Stock SQLite then reads the DDL, expects the
index, does not find it, and reports "database disk image is malformed
(11)" on any write or integrity_check.

So a table this crate creates with a declared composite key is not a
valid SQLite database, which is a direct hit on the crate's headline
byte-compatibility claim.

Recorded today in src/codegen/stmt/insert.rs's module doc:

A composite PRIMARY KEY(...)/UNIQUE(...) table constraint with no
backing CREATE INDEX/on-disk index (this codebase doesn't
auto-create sqlite_autoindex_* entries yet) has no real index to seek
against, so it still isn't enforced — that's a CREATE TABLE-side gap,
not an INSERT-codegen one.

#685 closed the INSERT-codegen half. This is the CREATE TABLE half.

Scope

src/codegen/ddl/ — CREATE TABLE must, for each constraint that
SQLite would give an autoindex:

• allocate an index b-tree root page;
• write a sqlite_master row with type = 'index',
  name = sqlite_autoindex_<table>_<n>, tbl_name = <table>, the new
  root page, and sql = NULL;
• number them by the rule fix: writes silently corrupt tables carrying a sqlite_autoindex_* (composite PRIMARY KEY / UNIQUE) #685 derived from the oracle and now
  implements in ddl_reader::autoindex_key_lists — declaration order,
  column-level constraints included, rowid-alias and WITHOUT ROWID
  primary keys skipped and consuming no number, redundant constraints
  collapsed.

The rule is already written and tested; this ticket consumes it rather
than re-deriving it.

Why it is separate from #685

Different blast radius. #685 touches the schema reader and refuses
unsafe writes; this touches DDL codegen and the freelist/page
allocator. #685 is what unblocks the SQE integration, because SQE's
warehouse files are created by C SQLite and only read and written by
this crate — it never creates them.

Complexity

Estimate: medium
Reasoning: The numbering rule and its tests already exist from
#685. The work is DDL codegen: allocating a root page per autoindex,
emitting the sqlite_master rows with a NULL sql, and keeping the
whole thing inside one transaction so a partial CREATE TABLE cannot
leave a table with some of its indexes. Populating the index is trivial
for a fresh empty table.

Acceptance Criteria

• A table created here with a composite PRIMARY KEY passes the
  oracle's PRAGMA integrity_check and accepts oracle writes
• No "database disk image is malformed (11)" from the oracle on any
  created shape
• sqlite_master rows match the oracle's byte-for-byte, including
  sql IS NULL
• Round trip: create here, write with the oracle, read back here
• Rowid-alias and WITHOUT ROWID tables gain no autoindex
• A failed CREATE TABLE leaves no partial index behind

Refs: 010/Req-8, #685