Channel size optimizations #573

Currently, we have no optimizations that affect the size of data travelling over a channel. In the presence of proc inlining such optimizations may cover some of the benefit of the optimizations in #561. In other multiproc implementation strategies, this can reduce the widths of FIFOs. Examples:

1. If the set of values sent over a given channel has size one, replace that channel with a single-bit channel.
2. Use range or ternary analysis to figure out the set of possible values sent over a given channel, and narrow it before sending and then unnarrow after receiving.
3. Use BDD analysis to find redundant bits in the values sent over the channel.
4. Find tagged union type data structures in the values sent over the channel. For example, if the channel has type (bits[1], bits[32], bits[64]) and when the bits[0] element is 0 the bits[64] is always 0, and when the bits[0] element is 1 the bits[64] is always 0, this can be reduced to (bits[1], bits[64]) where the bits[32] is sliced out of the bits[64] when it is nonzero. Simply supporting anonymous unions could be easier than doing this analysis, though it would only help in the case of the DSLX frontend, which does not support ADTs yet.

Doing this work interprocedurally before proc inlining has the benefit of doing less work overall, but the detriment of requiring more conservatism around any static analyses done.