tensor.insert kernel is broken into four cases:

• insertion indices are static/dynamic
• scalar v to insert is secret/cleartext

In all cases, this PR has the kernel generate a mask for the ciphertext-semantic scalar and dest tensors. The mask for the scalar has 1's or v's for the secret/cleartext cases, and the mask for the dest has 1's everywhere except 0's in the target of the insertion.

If the indices are static, then the insertions can be done directly: take the layout, add constraints for the static indices on the domain, then simplify and (using ISL) enumerate the points of the range. Each point becomes one insertion. A new utility getRangePoints was added to support this, which can be generalized as needed.

If the indices are dynamic, then we have to use the codegen, where the body of the loop does the relevant mask construction.

This improves over the previous insert kernel in that:

• Rotations are no longer required
• It supports dynamic insertion indices (previous one required static)
• It supports arbitrary layouts for the scalar (including multi-packed ciphertexts, fixing the tensor.insert parts of Support packing data across multiple ciphertexts in a tensor #1542 and ConvertToCiphertextSemantics: improve scalar layout handling and kernels for tensor.extract/insert #1662)

This one feels like another win for the new layout system, since it is so straightforward to express the assembly of a special plaintext mask.