add an implementation trading computation for memory

[clems4ever]

Realizing that the weights are also sparse, it might be interesting to exploit this fact to allocate only what is necessary in memory. This has the downside of being slower in terms query but I think this is definitely optimizable. This was just a proof of concept.

I took the opportunity to fix the other implementation and the test related to DyadicMemory.

[clems4ever]

I have made a little benchmark by trying to store all sequences of three consecutive uint8 values encoded with sdr size of 512 and population of 8. Each uint8 value is overlapping with the next one.
I can say that the memory gap is huge between this implementation and the performance optimized one, ~22k against 128M measured in number of actual stored values.

Obviously, the more there are values in the associative memory, the more consumed memory will tend to N^3 and the more it might impact performance. Again it is a tradeoff.

[clems4ever]

I have tried to use scipy sparse matrices (coo, csr, csc, dok) but they are way slower than defaultdict.

[PeterOvermann]

Thanks for your contributions!
One remark: About one third (1/e to be precise) of bits will be set when the memory is filled to capacity (n/p)^3 -- this is not sparse.