This reverts commit f6b50463.

Fear not, the roflscale will return when I have a bit more time and
figure out a better way to do it. (In particular, a way that doesn't
break the build.)

Don't prefix the name with attribute_. Avoids a string allocation on
read_attribute, which is a bit faster.

This means we never have to rely on define_method (which is slower and
uses more memory), even when we have attributes containing characters
that are not allowed in standard method names.

(I am mainly changing this because the duplication annoys me, though.)

Also make it actually work.

It slows down all read_attribute accesses to map 'id' to whatever the PK
actually is, inside read_attribute. So instead make sure the necessary
methods are defined and that they redirect wherever they need to go.

This is good because it reduces duplication.

This makes me happy!

Define singleton methods on the attributes module instead. This reduces
method pollution on the actual model classes. It also seems to make
something faster, I am unsure why! O_o

This reverts commit ee2be435.

If we don't have a primary key when we ask for it, it's better to fail
fast. Fixes GH #2307.

Return nil from read_attribute(:foo) if 'foo' is not present in the @attributes hash, but the _foo method has been defined. This brings the behaviour into line with the 3-0-stable branch and the master branch before 93641ed6 (there were previously no assertions about this which is why the change slipped through). Note that actually calling the 'foo' method will still raise an error if the attribute is not present.

Undo performances regressions I introduced in bbe0a507 and add test for an edge case. Add comments to explain the intent of the code.

Remove unnecessary code from define_read_method and add assertion to make sure the underscored version is actually generated