Implement linking optimizations

[msujew]

The linking phase is usually the most performance intensive phase. This is an attempt to improve its performance by reducing the amount of allocations performed during the linking.

The main improvement is concerned with the new SymbolSliceContainer. Adopters (or the generated <Language>SymbolContainer) can implement this to generate allocation free scopes.

Another major improvement is the reduction of per-reference allocations. Mainly the new context that we generate for linking. It's now a single context with a special resolutionChain attached to it that keeps track of cycles. Adds a ForEachWithSetup for this, which keeps one setup function per worker.

Then there are a few other improvements:

1. Reduce the amount of allocations in the parser by storing the closures for reference resolving functions
2. Store the name of a symbol on the symbol description directly. This reduces the need for interface dispatches (expensive).
3. Reduce the amount of closures required for traversing the AST.
4. Caches the available number of CPU cores. Calling GOMAXPROCS on every ForEach call had to acquire a lock first internally.

All of these changes together roughly double the performance of the workspace build.