performance sweet spot: balance between memory usage / CPU usage

[sbinet]

binary.Write creates a []byte slice corresponding to the whole input data (ie: for a []float64 slice, a make([]byte, len(slice)*8) is created and then filled).
this translates into better CPU performances (presumably because less calls to w.Write(buf[:])) but, of course, into a greater memory usage (and allocations).

not sure what npyio should do in this case (currently, it works element-wise, so just the memory to hold an element is used)
the CPU performances are not that bad.