weak references for "batched" profile compilation

[quattor-releaser]

One technique used at larger sites (MS, CERN) is to batch compilation requests (e.g. divide up into groups of 1000). Although this risks to have a few profiles recomputed, in reality, this rarely happens because cross-profile dependencies are rare. Essentially, this is the same as clearing the internal profile caches periodically. Determine if similar performance can be achieved by changing the internal profile/template caches to use weak references. This would then allow the normal garbage collection technique to clear out profiles that had not been used recently and reduce the memory requirements.

Reported by: loomisc