Change the shift apply in X and Y direction from 3/2 pixels to -3/2 pixels in
bundletoperfectsensor application for phr data. This
patch was provided by Julien Michel as he oberves that it improves the
fusion on PHR data.
I confirm that inverting this offset improve the result on PRIMARY product
samples over Melbourne.

It is still not clear for me if we've got now the "right" formula.

ENH: #942: rectify test so that leTvSOMClassifier and leTvSOMImageClassificationFilter can share the same baseline

Following recent commits on the ITK side which use clang-modernize to insert
missing override keyword, cf:

https://github.com/InsightSoftwareConsortium/ITK/commit/10757725be9cbe68ab2903c9cd4b76b738070e63

I had to compile OTB trunk using clang and  c++11 flags:

mkdir OTB-newclang
cd OTB-newclang/
CC=/usr/bin/clang CXX=/usr/bin/clang++ CXXFLAGS="-std=c++11 -stdlib=libstdc++" ccmake -DCMAKE_EXPORT_COMPILE_COMMANDS:BOOL=ON ../../../src/OTB

Internal libsiftfast does not compile as clang report error "error: argument evaluates to XXX
which cannot be narrowed to type YYY"

I added explicit cast to int in that cases.

Changes commited by mistake