Correctly estimate model parameters from warp fields

[nfaggian]

Another really important thing to get right is the initialization of the registration.

In many cases it is common to do a linear followed by a non-linear registration - so:

• perform linear registration - results in a warp field (tensor).
• estimate the spline coefficients that represent that warp field.
• start a non-linear registration using the estimated spline coefficients as a starting point for the optimization.

There needs to be "estimate" methods on all deformation models - so that a model can infer its parameters from a warp field.

[nfaggian]

This is related to ticket_26, it is handled by the estimate method on the CubicSpline class.