struct IR layout diverges from LLVM ABI layout

[resetius]

struct IR layout diverges from LLVM ABI layout

Status

• Fix agreed upon.
• Regression case added with {i32, f64} and an array of struct values.
• IR type table and LLVM backend use a single layout contract.
• Full Qumir regtest passes.

Scenario

For the struct {i32, f64}, the IR type table computes the size as the sum of
the field sizes and uses offsets 0 and 4. The LLVM backend, however,
represents the value as an ordinary unpacked LLVM structure {i32, double},
where double is located at offset 8 and the total size is 16.

When storing, accessing a field, or indexing an array of such structs, part of
the lowering uses packed IR offsets while LLVM load/store follows native
alignment. As a result, f64 is read starting at the wrong address and loses
its representation.

Expected behavior

Qumir should have a single, explicitly defined struct layout for IR, the VM,
and LLVM: field offsets, alignment, and final size must agree at every stage.
Another acceptable option is for the LLVM backend to use packed LLVM structs
if a packed layout is the IR's contract.

Workaround

generate explicit u8 padding fields