Before this change the usage of `strong_memoize`
resulted in using 4 gc slots (possibly with 4 mallocs)
for each call.

Depending on the data type a single memory allocation
is enough. This with improved implementation
results in 1.66x speedup.