TypeInType performance regressions

This ticket is to track the handful of performance regressions seen with the addition of TypeInType. It is quite possibly a good idea to break these out into separate tickets, but until we investigate, they're all bundled here.

The regressions are (all in bytes allocated, unless otherwise noted):

• T3064, up by 14.9%
• T5030, up by 61.8%
• T5837, up by 13%
• T5321Fun, up by 11%
• T5631, up by 39%
• T9872d, down by 22% (see below)
• T9872a, up by 33.6%
• T9872c, up by 59.4%
• T9872b, up by 49.4%
• T9675, up by 29.7%, and peak megabytes allocated up by 28.4%
• haddock.base, up by 12.4%
• haddock.Cabal, up by 9.5%

I did add an optimization around type family reduction (the function zonk_eq_types in !TcCanonical) that could cause such a drastic reduction.