There is more to come, but do a bunch now because it's cathartic.

When you grep the codebase for FIXME it's nice if the output isn't
overwhelmed by the same message repeated 20 times.

This matches the same variant in Terminal, matches the string
serialization "thresh", and will avoid weird compiler errors related to
collisions with the upcoming `Threshold` type.

This will be an annoying API-breaking change, but fortunately(?) this PR
is already breaking the API for thresholds in an annoying way.

In our threshold logic we have some special cases to handle multi,
multi_a, and conjunctions. However, we were not choosing between
these special cases correctly. Our logic was roughly that we would
try to use multi_a if all children were pk()s, OR try to use multi
if all children were pk() and there weren't too many, OR try to
use a conjunction if k == n.

The correct logic is: if all children are keys, and there aren't
too many, try to use multi or multi_a. ALSO, if k == n, try to
use a conjunction.

With this fix, the compiler correctly finds that conjunctions are more
efficient than CHECKMULTISIG when k == n and n < 3. When n == 3 the
two cases have equal cost, but it seems to prefer the conjunction. It
also correctly finds that when k == n, it is always more efficient to
use a conjunction than to use CHECKSIGADD.

This change necessitates changing some tests.