Fixes #1984

- Copy BETA2007.gsb, MD, alaska, conus, ntf_r93.gsb, ntv1_can.dat grids
  from proj-datumgrid to data/tests.
- Replace a couple uses of nzgd2kgrid0005.gsb in tests by ntf_r93.gsb
- Add downsampled/subsetted versions of egm96_15.gtx as tests/egm96_15_downsampled.gtx
  and ntv2_0.gsb as tests/ntv2_0_downsampled.gsb
  This results in a few changes in expected results
- Simpify travis/install.sh due to less configurations to test

This results in a hopefully acceptable increase of the proj-X.Y.Z.tar.gz
from 2.9 to 5.3 MB

Add alternative grid for JTSK - JTSK03 transform (EPSG:8364)

The definition of EPSG:8364 uses NADCON method for horizontal grid (.las/.los files)
but this format is not supported by PROJ.

UGKK (Slovak Geodetic and Cartographic Institute) provides NADCON .las/.los files here:
https://www.geoportal.sk/files/gz/slovakia_jtsk03_to_jtsk.zip

Additionally UGKK also provides the same grid file in NTv2 format:
https://www.geoportal.sk/files/gz/slovakia_jtsk03_to_jtsk_ntv2.zip

So let's add the NTv2 file the grid_alternatives table so that PROJ can automatically pick it up...

And add testing of both little-endian and big-endian NTv2 files

Fixes #1913

AuthorityFactory::createBetweenGeodeticCRSWithDatumBasedIntermediates() issued
a complex SQL query that pushes the SQLite3 query plan optimizer to its limits.
Was working reasonably with sqlite 3.11, but not with later versions.
So put less constraints in the main query and do post-processing checks and
auxiliary requests to avoid such issues.

For some unknown reason, this slightly slows down a bit execution time of the
whole test_cpp_api binary (~ 10%), but couldn't come with something better,
despite trying many variations of the main SQL query. It seems that in the
general case the non-filter LEFT JOIN on the supersession table helped,
except on this EPSG:7842 case.

Add EPSG records for 'Geocentric translation by Grid Interpolation (IGN)' (gr3df97a.txt) and map them to new +proj=xyzgridshift