Revisit memory allocation performance

Small I/O in the Hermes 2.0 platform is roughly 30% slower than the original Hermes 1.0 code in the single-thread case. This seems to be, in large part, due to the performance of memory allocation. Tasks have to be allocated from shared memory, which appears to account for this 30% difference. This can likely be reduced, as tasks are typically small (no more than 300 bytes).