[WIP] Remove Gaussian blur code alternatives that are exotic or didn't work very well

[griwodz]

Description

PopSift had a wide variety of downscaling and Gaussian filtering modes. These were written to increase the potential parallelism of PopSift, explore the trade-off between sequential operation and wider Gaussian filters, the impact of various filter width on quality, etc.

In practice, it seems that users stay with the default. The less successful downscaling/filtering options have therefore been removed.

Two options have been retained from the original PopSift code. Both implement the classical sequence that starts by scaling the input image and performing Gaussian filtering incrementally through the levels of an octave, and downscaling the following octave from the 3rd last level of the previous octave. These are the non-interpolating and the interpolating approaches.

The classical "non-interpolating" approach

It takes the neighbouring pixels and multiplies them with the weights from the pre-computed Gaussian filter. The width of this filter varies for each level to ensure a regular scale space.

The alternative "interpolating" approach

It starts with the same Gaussian filters, although their width is rounded up to a multiple of 2. The weight of two neighbouring cells g[n] and g[n+1] is then reformulated as a relative weight between these two cells, ie.:
g[n]I + g[n+1]J = abI + (1-a)bJ = b*(a*I + (1-a)J) where b=g[n]+g[n+1] and a=g[n]/(g[n]+g[n+1])

The appropriate values for b can then be read from an array, while the neighboring pixels I and J can be read with linear interpolation with a weight of A through the texture engine. This means that the term a*I + (1-a)J is handled by the texture hardware, leaving only one multiplication per 2 pixels.

Features list

• This PR does not add any features. It removes features.
• Faster and less accurate downscaling and Gaussian filtering modes have been removed.
• Enums for downscaling modes have been changed, "OpenCV mode" has been removed, "PopSift mode" and "VLFeat mode" have been renamed to reflect their actual functions.

Implementation remarks

• Removed fixed scaling code.
• Removed code to downscale everything directly from input image.
• Removed functions to interpolate from first image plane.
• Removed specialized version to create very first level from input image.
• Removed downscaling by interpolation, which could not be called with any parameter.
• Restructured the calling code for pyramid building to make sure that host code that starts the CUDA kernels is located in the same code file, making the CUDA kernels static. This prevents the overhead for linkable CUDA code.
• Simplified the solution with absolute sources. Returned to a solution without shuffle and identical code structure for horizontal and vertical Gaussian filtering.
• Simplified and unified code for absolute source interpolated Gaussian filtering.
• Renamed extrema refinement modes to have more intuitive names. They are no longer tied to PopSift vs VLFeat. (except that the command line parameters of the test code retains the old terms so far)

Handled in other PRs

• Remove Gauss filter tables for direct downscaling using absolute tables (done in PR Remove direct downscaling from input image to top level of every octave #178)
• Removed the config param ScalingMode (always use default) (done in PR Remove direct downscaling from input image to top level of every octave #178)
• Use horiz_from_input_image exclusively for octave 0. Direct downscaling is now only used for the input image. Note that initial blur is assumed for every input image, even when it is later interpreted as initially unblurred. That does make a difference, but is apparently recommended (done in PR Remove direct downscaling from input image to top level of every octave #178)
• Removed the narrower Gauss filter width that was called "OpenCV mode" (done in PR [sift] removed support for the old option for Gaussian filter width computation opencv #179)
• Removed deprecated scaling mode "OpenCV". OpenCV was buggy when this code was written. It has improved since then. (done in PR [sift] Remove unreachable code #180)

[griwodz]

PR #162 has been merged because latest CUDA for Linux can no longer run without it.