Fix segfault when matching tuple elements against unions or interfaces via Any

[SeanTAllen]

When pattern matching a boxed tuple (via Any val), unboxed numeric elements were loaded as pointers when the match pattern expected a union or interface type. The raw byte value got interpreted as a memory address and the program segfaulted.

The codegen for dynamic_capture_ptr and dynamic_value_ptr now checks the runtime type_id to distinguish object pointers (odd type_id, load directly) from raw inline data (even type_id, box into a heap object first). The boxing helper allocates via pony_alloc, stores the descriptor, and memcpys the data at the correct alignment-aware offset. This uses the same type_id encoding that gentrace.c already relies on.

The fix also corrects make_field_offset to return the actual struct offset for boxable numeric types. It was returning 0 for all types with field_count == 0, which would put the value at the wrong offset for types like I128/U128 where alignment padding pushes the value past the pointer-sized offset.

Five regression tests cover: union-typed element, interface-typed element, the no-double-boxing scenario from PR #4787's regression, I128 alignment, and nested inline tuples.

Note: The dynamic_value_ptr fix mirrors dynamic_capture_ptr but has no dedicated test. I think the pointer-type branch there is probably unreachable in practice since value patterns resolve to concrete eq() parameter types rather than unions or interfaces. The code is defensive.

Closes #4507

[SeanTAllen]

@nisanharamati please give this a test.