Fix alignment in relaxed atomics spec test (#8961)

The spec test intends to use the natural alignment for all instructions,
but it's choosing an alignment based on the memory address type rather
than the size of the load/store. This means that sometimes the alignment
is too small (4 bytes for an 8-byte load/store in a 32-bit memory), and
sometimes the alignment is too large (8 bytes for a 4-byte load/store in
a 64-bit memory) both of which are a [validation
error](https://github.com/stevenfontanella/threads/blob/f521d7b3c19bcc9097aad979b3b91837eab88d83/interpreter/valid/valid.ml#L206)
(but are not in Binaryen today).

Use the size of the load/store as the alignment instead.

Spec reference:
<img width="737" height="216" alt="image"
src="https://github.com/user-attachments/assets/b86de03e-1316-4c2c-9a12-80773fb88119"
/>

Author: stevenfontanella

scripts/test/generate_atomic_spec_test.py

@@ -224,7 +224,7 @@ def bin_statement_lines(template: Template, mem_idx: int, mem_ptr_type: ValueTyp
 
     has_ordering = ordering is not None
     has_mem_idx = mem_idx is not None
-    raw_alignment = int(math.log2(mem_ptr_type.value // 8))
+    raw_alignment = int(math.log2(template.value_type.value // 8))
     alignment = raw_alignment | (has_ordering << 4) | (has_mem_idx << 6)
     comment = f"Alignment of {raw_alignment}" \
               f'{" with bit 4 set indicating that an ordering immediate follows" if has_ordering else ""}' \