Increase sparsity of pippenger fixed window naf representation

[jonasnick]

Fixes #506

[jonasnick]

ecmult_multi pippenger speedup

without endo:    2 points 11%,     pippenger threshold (160 points) 0.74%
with endo:       2 points 10%,     pippenger threshold (88 points)  2.1%    

[jonasnick]

Added wnaf_fixed bugfix and two commits that further increase sparsity. Now it's possible to use precomputation with pippenger similar to using the endomorphism: split a scalar, point pair (s, G) into (s1, ..., sn, G, ..., Gn) where all si have 256 -256/n leading zeros and s1*G + ... + sn*Gn = s*G.

With 128 points and n=8 this gives a speedup of 6% over endomorphism (13% over non-endo). See https://github.com/jonasnick/secp256k1/blob/bench_ecmult_precomp/src/bench_ecmult_precomp.c#L138 for the benchmarking and splitting code.

[sipa]

Lightly tested ACK ec0a7b3

[apoelstra]

Let me test it with the BP code before merging.

[apoelstra]

ACK