fix(sqlite): don't dereference a NULL declared type in column_nullable

[msdrigg]

Fixes #4373

sqlite3_table_column_metadata() leaves its declared-type out-param NULL when a column was declared with no type (CREATE TABLE foo (bar PRIMARY KEY) is legal SQLite). column_nullable() passed that pointer straight into CStr::from_ptr(), so describing such a column segfaulted. Since query!() describes a live database while expanding, the process that died was rustc.

Introduced in 69ee0df (#4088), released in 0.9.0. Before that the slot was ptr::null_mut() and never read.

The fix. Only read the pointer when it's non-null. A column with no declared type isn't declared INTEGER, so it can't be a rowid alias, and it falls through to the ordinary NOT NULL check:

let is_integer = !datatype.is_null()
    && CStr::from_ptr(datatype)
        .to_bytes()
        .eq_ignore_ascii_case("integer".as_bytes());

Ok(if primary_key != 0 && is_integer {
    None
} else {
    Some(not_null == 0)
})

This matches what column_decltype() right above already does with the pointer from sqlite3_column_decltype().

That's also the correct answer semantically, not just a crash guard: a PRIMARY KEY with no declared type has BLOB affinity, is not a rowid alias, and SQLite really does allow NULLs in it, so Some(true) is right.

After the fix, an untyped column gives the normal diagnostic instead of a crash:

error: no built-in mapping found for type NULL of column #1 ("device_id");
       a type override may be required

and a type override (device_id as "device_id!: String") compiles cleanly. That's the behaviour #1979 and #3546 describe.