memory optimization for aggregator

[KiKoS0]

Bunch of optimizations regarding memory allocation, reducing lock contention in pauses aggregator:

• MatchResult was flattened from a nested map to a single map[matchKey]int, removed mutex from it so it's no longer thread safe but the synchronization overhead of multiple goroutines writing to it likely outweighed the parallelism gains.
• Add a sync.Pool for MatchResult to reduce allocations per eval.
• Removed goroutine pools from engine Match() methods, the spawning/scheduling/atomics overhead was likely beating the purpose of actually parallelizing these typically small iterations.
• Remove() is now O(1): just stores the ID of the deleted evaluable in a sync.Map. A background gcWorker goroutine periodically drains the deleted set in batches so that we never block evaluations for long periods. Engine Remove() also takes a batch now, single lock acquisition per batch.
• rand.Rand can also be pooled via sync.Pool in the parser.

[darwin67]

nit: probably can delete these lines now