We'd previously been assuming we'd want to predict P-256 and X25519 but,
on reflection, that's nonsense. Although, today, P-256 is widespread and
X25519 is less so, that's not the right question to ask. Those servers
are all 1.2.

The right question is whether we believe enough servers will get to TLS
1.3 before X25519 to justify wasting 64 bytes on all other connections.
Given that OpenSSL has already shipped X25519 and Microsoft was doing
interop testing on X25519 around when we were shipping it, I think the
answer is no.

Moreover, if we are wrong, it will be easier to go from predicting one
group to two rather than the inverse (provided we send a fake one with
GREASE). I anticipate prediction-miss HelloRetryRequest logic across the
TLS/TCP ecosystem will be largely untested (no one wants to pay an RTT),
so taking a group out of the predicted set will likely be a risky
operation.

Only predicting one group also makes things a bit simpler. I haven't
done this here, but we'll be able to fold the 1.2 and 1.3 ecdh_ctx's
together, even.

Change-Id: Ie7e42d3105aca48eb9d97e2e05a16c5379aa66a3
Reviewed-on: https://boringssl-review.googlesource.com/10960
Reviewed-by: David Benjamin <davidben@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
CQ-Verified: CQ bot account: commit-bot@chromium.org <commit-bot@chromium.org>