Rewrite the eclass-reading code

[pavpanchekha]

This PR basically attempts to optimize split-untyped-eclasses, which on its own is about 7.5 minutes of runtime in a nightly. To do so it:

• Uses sequences of "stream in" enodes instead of ever allocating a vector of all the ones in an eclass. This also reduces allocations.
• Micro-optimizes the transformations from types and egg eclass ids to Herbie-side ids
• Unpacks enodes in a couple of cases to save a minute number of allocations

The results code is fairly ugly (it is almost C code) and it's not clear if the result is actually much faster. Life is hard!

[pavpanchekha]

Not faster (slower)