Array reinterpret operations

[mshinwell]

This is a different approach from the one I took previously to adding the "array reinterpret" operations, e.g. given an int64# array, get the two values corresponding to the type #(float# * float#) from byte offset 8 (which is unaligned with respect to the product type as a whole). I think this way is easier: the Lambda primitives now mirror what happens in the middle end, namely they take both the array kind and the array ref/set kind. (This means that all this information basically just passes through Lambda_to_flambda, which in fact could now become simpler, as some conversion functions could now be deleted were it not for the float array optimization.)

• The array kind determines indexing
• The array ref/set kind determines what is actually read or written.

We should be able to expose reinterpret primitives in Translprim which map to these existing primitives, just with the kinds appropriately set, I hope.