Add address_type=auto

[akiradeveloper]

Add support for #[cube(..., address_type = "auto")].

Automatic mode selects the concrete AddressType at launch time from the
largest indexable buffer length used by the kernel:

• U32 when every buffer length fits in u32
• U64 when at least one buffer exceeds u32::MAX

The resolved concrete address type is included in the existing kernel settings
and cache key.