Fix inconsistency: expected scale not be subject to scale influence

We should treat all three "target" ("expected") parameters in a similar way:
The "influence" control should only work on the measurement part of stabilisation,
i.e. it should only control the automatic part of stabilisation, while
the target parameters are deliberately set by the user and thus should
even be in effect when the automatic stabilsation is turned down.

It used to be so for location and rotation, but for the scale part,
I re-used the existing code for autoscale, which also had the scale influence
work on the autoscale factor. This was sensible in the old version,
since scale_influence was the only way to control the result. But now,
the user has always total control trough the "target_*" parameters
and thus we should prefer to treat all similar.

2D stabilization: flip orientation of the scale parameter

values > 1 will zoom in and values < 1 zoom out

Rationale: the changed orientation is more natural
from a user POV and doing it this way is also more
consistent with the calculation of the other
target_* parameters.

Compatibility: This will break *.blend files saved
with the previous version of this patch from the
last days (test period). It will *not* break any
old/migrated files: Previously, the DNA field "scale"
was only used to cache autoscale. Only with the
Stabilisator rework, "scale" becomes a first class
persistent DNA field. There is migration code to
init this field to 1.0

2D stabilization: change presentation of target_scale in UI

Alongside with this change, we fix disabling of Autoscale,
because this feature works even when rotation/scale is disabled.

2D stabilization: by default init anchor_frame to frame 1

It is common in blender to use 1-based counting for
frame sequences (while 0-based is allowed). Thus
initializing to use frame 1 as reference for stabilization
is likely to produce smooth start values in most cases

Change Request: use weight centre of location tracks as pivot

Previously, this extension used the translation compensated image centre
as reference point for rotation measurement and compensation. During
user tests, it turned out that this setup tends to give poor results
with very simple track configurations.

This can be improved by useiing the weighted average of the location
tracks for each frame as pivot point. But there is a technical problem:
the existing public API functions do not allow to pass the pivot point
for each frame alongside with the stabilisation data. Thus this
change implements a trick to package a compensation shift into
the translation offset, so the rotation can be performed around
a fixed point (center of frame). The compensation shift will then shift
the image as if it had been rotated around the desired pivot point.

Rework 2D stabilizator

See this page for motivation and description of concepts:
https://github.com/Ichthyostega/blender/wiki

See this video for UI explanation and demonstration of usage
http://vimeo.com/blenderHack/stabilizerdemo

This proposal attempts to improve usability of Blender's image stabilization
feature for real-world footage esp. with moving and panning camera. It builds
upon the feature tracking to get a measurement of 2D image movement.

- Use a weighted average of movement contributions (instead of a median).
- Allow for rotation compensation and zoom (image scale) compensation.
- Allow to pick a different set of tracks for translation and for
rotation/zoom.
- Treat translation / rotation / zoom contributions systematically in a
similar way.
- Improve handling of partial tracking data with gaps and varying
start / end points.
- Have a user definable anchor frame and interpolate / extrapolate data to
avoid jumping back to "neutral" position when no tracking data is available.
- Support for travelling and panning shots by including an //intended//
position/rotation/zoom ("target position"). The idea is for these parameters
to be //animated// by the user, in order to supply an smooth, intended
camera movement. This way, we can keep the image content roughly in frame
even when moving completely away from the initial view.

A known shortcoming is that the pivot point for rotation compensation is set to
the translation compensated image center. This can produce spurious rotation on
travelling shots, which needs to be compensated manually (by animating the
target rotation parameter). There are several possible ways to address that
problem, yet all of them are considered beyond the scope of this improvement
proposal for now.

Own modifications:

- Restrict line length, it's really handy for split-view editing
- In motion tracking we prefer fully human-readable comments, meaning we
don't use doxygen with it's weird markup and comments are supposed to
start with capital and end with a full stop,
- Add explicit comparison of pointer to NULL.

Reviewers: sergey

Subscribers: kusi, kdawg, forest-house, mardy, Samoth, plasmasolutions, willolis, sebastian_k, hype, enetheru, sunboy, jta, leon_cheung

Maniphest Tasks: T49036

Differential Revision: https://developer.blender.org/D583