ENH: Potential avoidable overhead in Lasso/celer_path

[mathurinm]

Thanks to @josephsalmon

Celer first point is around 0.8 s and then convergence is fast. it seems like some first computation takes a long time and Blitz does not suffer from this issue.

Profiling may help if the issue comes from python instead of cython