Make sure that input memref to `upmem.scatter` is legal, that is, all scattered elements must be contiguous in memory

[oowekyala]

This is in part enforced by the CNM scatter map shape. The scatter map mandates that the last dimensions of the input have the same shape as the buffer. However since then, CNM scatter has been made bufferizable. If the memref has a custom layout or strides, the elements may not be contiguous anyway.

For instance

%3 = scf.for %arg6 = %c0 to %c1024 step %c32 iter_args(%arg7 = %alloc_0) -> (memref<1x128xi32>) {
          %subview_2 = memref.subview %arg0[%arg2, %arg6] [1, 32] [1, 1] : memref<1x1024xi32> to memref<1x32xi32, strided<[1024, 1], offset: ?>>
          %4 = upmem.alloc_dpus : !upmem.hierarchy<1x128x1>
          upmem.scatter %subview_2[132, 32, #map] onto %4 : memref<1x32xi32, strided<[1024, 1], offset: ?>> onto !upmem.hierarchy<1x128x1>

is this legal?
-> yes, because the stride in the second dimension is 1.

[oowekyala]

This can currently only be verified at the CNM level. The thing is not the entire input memref is required to be contiguous, only the dimensions that we are transferring (which is the shape of the cnm.buffer). However at the upmem level, the cnm.buffer has disappeared, and the scatter/gather ops are operating on the whole memref with a linear index. We have lost information about which dimensions need to be contiguous, and which don't. We could add an attribute to upmem.gather/scatter to be able to verify that though.