discretization: fix logic to detect if polynomial storage should be resized

[andrea-iob]

The current storage size was not properly evaluated.

I've also updated the function that evaluates the polynomial coefficients: since the weight matrix is usually small, it's faster perform matrix vector multiplication using an ad-hoc code rather than using CBLAS functions..

[andrea-iob]

I've put back the CBLAS function in the evaluation of the polynomials. The raeson I was seeing poor performances for small matrices may be related to the fact that I was using an unoptimized BLAS library.

[andrea-iob]

Rebased on current master.