upb: reject aligned MiniTable size overflow

[heap-fixer]

This patch fixes a MiniTable size alignment overflow in upb/mini_descriptor/decode.c.

Previously, MiniTable construction checked the unaligned message layout size against UINT16_MAX, stored it in the 16-bit upb_MiniTable::size field, and only then aligned it up to kUpb_Message_Align. For sizes near UINT16_MAX, that alignment step could produce 65536, which truncated to 0 when written back to the 16-bit size field.

That could leave a MiniTable with:

• size == 0
• non-zero field offsets

and later cause _upb_Message_New() / upb_Message_New() to allocate a zero-sized arena object for a non-empty message layout.

This patch fixes the issue by:

• computing the aligned size in size_t
• checking the aligned result against UINT16_MAX
• rejecting oversized MiniTables before truncation
• only storing the value into upb_MiniTable::size after validation

It also adds a regression test covering the alignment-overflow boundary, so MiniTable construction now fails instead of producing a wrapped zero-sized layout.