Precise bounds calculation for nfixed_mat_mul

[fredrik-johansson]

For each of

• _nfixed_mat_mul
• _nfixed_mat_mul_classical
• _nfixed_mat_mul_waksman
• _nfixed_mat_mul_strassen
• _nfixed_complex_mat_mul

we implement a corresponding method that bounds the magnitudes of all intermediate results as well as the final error (in ulps).

The magnitude bound allows one to scale the inputs tightly while guaranteeing |x| < 1 for intermediate results so that no overflow is possible.

Also, thanks to the error bound, it will be possible to use nfixed methods for arb_t and acb_t matrix multiplication (not done in this PR).

@albinahlback This basically SLP error analysis, so you might want to give this code a careful look when you have time.