Optimizing the memory pipeline

[winston-h-zhang]

This is an umbrella issue for optimizing the memory allocations in arecibo. The original whole PR is contained in #118 -- however, to have a higher degree of confidence in the breakdown of optimizations, we split the changes into several self-contained PRs. For each of these separate PRs, we provide benchmarks to show that they individually contribute to better performance.

Tracking PRs

• Allocation free synthesis #138, replaced by Preallocate witness before synthesis #143
• Reduce allocations/cloning in spmvm #139
• Remove large vector allocations #144; depends on Allocation free synthesis #138 Preallocate witness before synthesis #143, Reduce allocations/cloning in spmvm #139

TODOs

• Port changes to SuperNova (done in Cleanup memory pipeline for SuperNova #227)
• Port changes to ParaNova

[huitseeker]

Note : those changes should be ported to Supernova relatively quickly, as #197 threatens to replace the old (memory-optimized) code paths with the batched (non-memory-optimized) ones, losing the progress we've made here.

[huitseeker]

@winston-h-zhang Can we close this after #227 ?