cuda.core: grown VMM buffers record the requested size, not the aligned size

modify_allocation records new_size on the returned Buffer (slow path: Buffer.from_handle(ptr=new_ptr, size=new_size); fast path: buf._size = new_size), while allocate() records the granularity-aligned size. For a grow to a size that is not a multiple of the granularity (2 MiB on current GPUs), close() calls cuMemUnmap with an end inside the last mapping, which the driver rejects, so the reservation, the mappings, and the physical memory leak and the Buffer deleter reports a CUDAWarning. The equal-size branch hits the same alignment rule in cuMemSetAccess and raises. Reachable today through the slow path. No test grows to an unaligned size. Fix belongs to the ownership redesign (epic #2906).

Start in modify_allocation and compare the slow path's Buffer.from_handle size with the fast path's buf._size and allocate()'s aligned size. Exercise an unaligned grow and the equal-size branch, then verify close() reaches cuMemUnmap and cuMemSetAccess without CUDA warnings or leaks. Account for the ownership redesign in epic #2906 and add regression coverage for these cases.