Implement dead code elimination/value numbering for shuffles

[rachitnigam]

We should run DCE after register allocation and shuffle truncation. We generate quite a few unnecessary registers.

[avanhatt]

This would be good to do for human-readability of the kernels Vincent is generating (mostly in removing the statically specified shuffle arrays that are rendered dead).

[avanhatt]

There are a few different places in the code this could slot in:

Currently, shuffle-truncation-pass.rkt is where these shuffle indices are created, so we could handle deduplication there. Alternatively, we could add a new dead-code-elminiation-pass.rkt file to do just the deduplication and dead code elimination passes there. That pass would get called from compiler.rkt

[avanhatt]

One path forward: starting in shuffle-truncation-pass.rkt, build a map of final are already seen, and reference an existing shuffle instead of building a new duplicate one where possible. Then have a second pass to make sure any other "dead" shuffles are removed (likely working through the list of instructions backwards, removing any shuffle declaration where you haven't seen a use).