Variable time normalize #137

This may not be worth it; I measure around a 0.8% speedup for verification.

I'm inclined to take even 'small' improvements like this as they'll have a greater relative impact as other things are improved. I still need to test and reason about this some more.

Adding instrumentation to check the normalized debugging variable right now tells me that 10% of the calls to normalize during a bench_verify call were provably already normalized (assuming the instrumentation isn't wrong).

This seemed awful high to me, so I instrumented it to find out what the backtrace was when it was already normalized. Turns out that in secp256k1_gej_add_ge_var at group_impl.h:305 it's always normalized, well, no shock:

    secp256k1_fe_t u1 = a->x; secp256k1_fe_normalize_var(&u1);
[... no access to u1 ...]
    secp256k1_fe_normalize_var(&u1);

This is responsible for 99.72% of the already verified calls. I created #139 for this. I'm doubting that it will make a huge benchmark difference, but it might be useful to check.

ACK.