Switch evaluation environment from Data.Map to Data.IntMap.

[brianhuffman]

Benchmarks using examples/AES.cry and Primitive/Keyless/Hash/SHA1.cry from the cryptol-examples repo show that using IntMap instead of Map for the value environment yields a speedup of about 7 or 8 percent, and a reduction in memory footprint of about 4 or 5 percent.

[robdockins]

We're already doing the working of computing unique names, so this seems like a pretty straightforward way to squeeze out some performance. I'm looking into doing basically the same for the types environment (type evaluation consumes a surprising amount of time for some workloads).