[SCFToCalyx] Modify top-level function in place and propagate external memory allocations

[jiahanxie353]

This patch refactors the logic when there are memref type function arguments in the top-level function (because Calyx only supports interacting with memory data if that memory is allocated using @external and does not support function argument memory data passing).
We used to create a new top-level function, which does nothing but invokeing the old top-level function. This approach is fine semantically, but recently @ayakayorihiro discovered that this prevents the static inference from inferencing the number of cycles in the Calyx project: calyxir/calyx#2452.
As a solution, we can just modify the top-level function in-place by moving those memref function arguments to explicit memref.allocs in the function body.

Moreover, this patch also supports propagating external memory allocations from non-top-level functions to the top-level function because of the constraint in the Calyx project: https://github.com/calyxir/calyx/blob/50f817f9991df4fa87f2178222ba45a026aba549/calyx/opt/src/passes/well_formed.rs#L185