When converting an image TRC (or a precision+TRC change), we used to
convert the profile to a linear (resp. sRGB TRC) variant or use the
builtin profile, which we'd set on the image.

Pippin is telling me that "it used to be that linear and non-linear
precision gave different results, rather than only about how things were
stored". Yet this is not needed anymore. Nowadays "RGBA foo" + any
profile is linear anyway, disregarding the profile's TRC.

Furthermore, this way, we don't lose the original TRC in a non-linear ->
linear -> non-linear roundtrip and we simplify the code, which fixes
such problems as shown in #3495 as a side effect.