To allow MatrixDouble to be a drop-in replacement for gfxMatrix, it
needs to accept the "double" versions of Point, Rect, and Size. This
patch does that by adding some extra typedefs inside BaseMatrix to
abstract over that.

It also moves some function implementations into the .h file as they
don't need specialization. I left some function implementations in
the Matrix.cpp file:
- Rotation, because it is specialized for Float and Double, since it uses
  sinf/cosf vs sin/cos in the two implementations.
- The Matrix4x4 multiplication operator overload, because if I put it
  inside the BaseMatrix class declaration Matrix4x4 isn't defined yet
  and the compiler doesn't like it.

MozReview-Commit-ID: K56dZjJhXWS

--HG--
extra : rebase_source : d009151942811a725ee6bf854238073e59665d5c