A slightly faster slimmer curl

I feel like the next logical step would be to allow callers to pass their own allocation functions. I think that moving from many small allocations to one large one - like with a memory arena - would perhaps improve the performance if moving from 100 -> 96 did.