perf: use width-8 NAF for static scalars to match NafLookupTable8

[GarmashAlex]

The precomputed Straus implementations for both the vector and serial backends build NafLookupTable8 for static points, which stores odd multiples up to 127A, but they decomposed static scalars with a width-5 NAF, restricting lookups to 1..15 and leaving most of the table unused. This change switches static scalar decomposition to non_adjacent_form(8), fully utilizing the 64-entry odd table. It reduces the number of static additions due to the lower nonzero digit density of width-8 NAF while keeping the number of doublings and the memory footprint unchanged, since the larger table was already computed. The change is safe because Scalar::non_adjacent_form supports w<=8 and NafLookupTable8::select requires odd digits less than 128, and it matches the existing pattern in vartime double-base multiplication where width 8 is used when a large odd-multiples table is available. No API or behavioral changes beyond this decomposition width are introduced.