Some more documentation for in the rasterizer

[nical]

Some of it requires a bit more attention than the previous PR.

• The value stored in B is multiplied by 2 (so that it matches the derivative formula at t=0), which is compensated by not multplying B by two later when it is used.
• Some code was written as foo * (1 << (16 - SAMPLE_SHIFT) got reformulated as dot2_to_dot16(foo). I didn't do it initially because I wasn't sure whether the doing foo * 1 << bar instead of simply foo << bar was indicative of something that I had missed but I think it makes better sense in the new formulation.
• I removed the comment saying "we'll iterate t from 0..1 (0-256)". As far as I can tell we aren't manipulating the curve parameter t anywhere. This comment had me puzzled for a while, if it indeed conveys something meaningful I think it needs to be rephrased.

[nical]

Note: the skia equivalent looks simpler: https://github.com/RazrFalcon/tiny-skia/blob/afa3172310bb7a30e44911ef825f06dd900a4d85/src/edge.rs#L234 (note the comment above about preserving a bit by not multiplying by two and compensating in the curve shift).