Profile and optimize frontend (from CGF submission to serialized graph)

[PeterTh]

#107 has frontend benchmarks that do not instantiate the entire runtime.

Low hanging fruit:

• Currently, 13% of the time is spent in iostreams for creating the CDAG debug labels. Maybe this should happen in graph printing, not generation (there is already a PoC implementation for that)
• malloc/free pairs have significant impact deep in dependency checking. We can potentially eliminate a lot of convenience allocations (get_accessed_buffers et al)

Requires further investigation:

• Use a bump allocator for intrusive graphs to improve locality, e.g. a heuristically sized pool per horizon + fallback as needed (maybe also for the dependencies vectors?)
• command / task map rehashing is also a noticeable factor, but memory is abundant - reserve() them to avoid stalls

[PeterTh]

I think this is actually (reasonably) complete at this point. The low hanging fruit has been addressed by various changes and I don't think we are substantially limited by "frontend" performance in real-world benchmarks right now.