Improvements

[romainsacchi]

• Faster approach to calculating the results: does not longer calculate the LCA for each car, but rather calculates the LCA for each first level activity of all cars and then multiplies by the required quantities to obtain the score of all cars, using broadcasting. Also uses numexpr to do some of that.
• This approach also allows to better allocate impacts between categories. Hence, there is no longer the "other" category, as we now know how the residual should be allocated. Instead, we have a new "End of life" category.
• Downgraded some of the arrays to float16, to gain memory. Same for the array containing the results.