[DRAFT] Use arena allocator when exporting goto binaries

[AlexanderPortland]

Motivation

A large portion (~7%) of our codegen time is spent simply deallocating the large SymbolTable struct that we generate when building each goto binary for CBMC. This ends up taking a while because the struct contains Irep structs which can each recursively contain multiple Vecs of even more Ireps, requiring a lot of small allocations to be individually handled over and over again.

Approach

This PR switches all data structures within Kani's binary generation to use a shared bump allocator (using the Bumpalo crate) that ensures all allocations we use while serializing a given binary remain in a single large region of memory. Once a given binary has been serialized, we can then easily deallocate all the memory it used at once.

As part of this change, we also replace the LinearMap used to represent named subs within an Irep with a proper hashbrown::HashMap.

Issues

Based on manual testing there is currently a small memory leak (I think because the underlying vectors for the BigInt & BigUInt types we use here) still use the global allocator and thus will not be deallocated with the memory arena. I'm looking into fixing this and improving code comments for the more involved changes, but wanted to make a draft PR to get preliminary performance results first.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 and MIT licenses.