Merge #1316: Do not invoke fe_is_zero on failed set_b32_limit

6433175 Do not invoke fe_is_zero on failed set_b32_limit (Pieter Wuille)

Pull request description:

  Noticed in the CI output of #1313 (https://cirrus-ci.com/task/5117786435878912)

  The code violates the field element contract that states that a field element that comes out of a failed `secp256k1_fe_set_b32_limit` call cannot be used before overwriting it. This is not an issue in practice, as such failure can only occur with negligible probability, but the experimental compiler in that CI setting is technically correct in detecting this possibility.

  Fix it by setting it to 1 based on a `secp256k1_fe_normalizes_to_zero` test rather than a `secp256k1_fe_is_zero` one (which does not require normalization).

ACKs for top commit:
  stratospher:
    ACK 6433175
  real-or-random:
    utACK 6433175

Tree-SHA512: 49da4535181c4607c1f4d23d1fd7cd65e7751c7cfa68643f1da77f3ec7961754fc8553bb415137fd61d86c805fe69f5adf97c05b9dc4d3bf357ae7c6409cc51a

Author: real-or-random

src/ecmult_gen_impl.h

@@ -87,7 +87,6 @@ static void secp256k1_ecmult_gen_blind(secp256k1_ecmult_gen_context *ctx, const
     secp256k1_fe s;
     unsigned char nonce32[32];
     secp256k1_rfc6979_hmac_sha256 rng;
-    int overflow;
     unsigned char keydata[64];
     if (seed32 == NULL) {
         /* When seed is NULL, reset the initial point and blinding value. */
@@ -106,11 +105,9 @@ static void secp256k1_ecmult_gen_blind(secp256k1_ecmult_gen_context *ctx, const
     memcpy(keydata + 32, seed32, 32);
     secp256k1_rfc6979_hmac_sha256_initialize(&rng, keydata, 64);
     memset(keydata, 0, sizeof(keydata));
-    /* Accept unobservably small non-uniformity. */
     secp256k1_rfc6979_hmac_sha256_generate(&rng, nonce32, 32);
-    overflow = !secp256k1_fe_set_b32_limit(&s, nonce32);
-    overflow |= secp256k1_fe_is_zero(&s);
-    secp256k1_fe_cmov(&s, &secp256k1_fe_one, overflow);
+    secp256k1_fe_set_b32_mod(&s, nonce32);
+    secp256k1_fe_cmov(&s, &secp256k1_fe_one, secp256k1_fe_normalizes_to_zero(&s));
     /* Randomize the projection to defend against multiplier sidechannels.
        Do this before our own call to secp256k1_ecmult_gen below. */
     secp256k1_gej_rescale(&ctx->initial, &s);