During this #6131 (comment) @bkchr discovered that the inline_size of the SharedCache balloons to 3GiB when running state_access benchmark and to 1.5GiB if you run the read/write storage benchmarks with asset-hub-westend.

This happens because the size_of of NodeOwned structure is 750 bytes, and when it is used here

The NodeOwned is that big because of this enum variant, which uses a static array of size 16.

	NibbledBranch(
		NibbleVec,
		[Option<NodeHandleOwned<H>>; nibble_ops::NIBBLE_LENGTH],
		Option<ValueOwned<H>>,
	),

The tests ran on state_access benchmarks and on asset-hub-westend showed that we could reduce the used space by a factor of 4 if the array would be dynamically allocated only for the populated children.

We need to be careful here not to fragment the memory with so many small allocation, see #6131 (comment), one idea @bkchr had is to use an arena allocator.