The following will take much less time and can achieve good, native results quickly:

Design a user-code facing clean api using idiomatic D (front end code): windows, widgets, callbacks via delegates, etc.
Design a glue layer to talk to different backends: gtkd, wxd, qtd, fltk etc.

This is what python does with matplotlib:
http://matplotlib.org/faq/usage_faq.html : they support pygtk, wxpython, tkinter, qt, macosx, or fltk, and also non interactive backends)

The user code stays clean, the results are native (depending on backend), and the wheel is not reimplemented.

The other advantage of this is we can have results early on (leveraging existing backends), but nothing prevents us to work in parallel on a backend that's purely written in D (written on top of opengl for example) if need be.