Introducing the Horner polynomial evaluator also introduces the need for
very long +init:tag arguments (a n'th order 2D polynomium has
(n+1)(n+2)/2 coefficients, and n is typically in the range 5-10, i.e. up
to around 60 coefficients for each polynomium, and there are 4 polynomia
in a complete back/forward transformation set).

Hence, in this commit, along with the first part of the Horner code, the
code for reading +init files has been modified in a (for all practical
purposes) backwards compatible way, by making it possible to introduce
line continuations by escaping line breaks, i.e. preceding them with a
backslash.

An escaped line break works (as it would in TeX), by skipping all
following whitespace, including interspersed #-comments. This simple
extension makes it possible to create very long initialization elements
without losing track of the structure (cf. s45b.pol and pj_init_test.c
in the examples-directory for a demo).

The s45b.pol file was created by hand-editing the output of the software
doing the original constrained adjustment for the polynomial
coefficients. The simple adding of the “skip following whitespace and
comments” feature has made it possible to retain almost all metadata
from the source material.

This is considered very important, since 1) For the lack of a prior
common file format for geodetic polynomial coefficients, there is a good
chance that this will become THE standard, at least for the time being,
and 2) Without the metadata represented, it will be very hard for a
human to debug code involving a slightly misrepresented polynomium.

Due to the current architecture of the pj_init.c code (mostly around the
fill_buffer() function), it is next to impossible to implement the line
continuation functionality in full generality. Hence, it has been
necessary to limit this format extension to files smaller than 64 kB.

Getting ready for some consultancy with Karsten Engsager, who has
maintained the original code for more than 2 decades.

The line continuation format extension was misimplemented: The comment
skipping logic was placed on the wrong side of a while loop

The first HEALpix test case in nad/testvarious is clearly intended to
invoke the spherical form of HEALpix.

It does, however, specify the spheroid using the +a=1 size parameter,
without specifying any shape parameter.

But since +no_defs is not specified either, a shape parameter is picked
up from the nad/proj_def.dat file (where ellps=WGS84 is given in the
<general> section).

For reasons that I fail to comprehend, it appears that this has not
happened before I updated the pj_init code to support projection
pipelines.

I do, however, believe that the present behaviour is the correct one,
and rather than retrohacking the pj_init code, to (incorrectly, I
believe) reproduce the prior behaviour, I have corrected the test case
invocation in nad/testvarious to specify the spheroid using the +R=1
size parameter (which was already used in the following test case).

<horner.h> appears to work with  cmake build, but not with autotools

(Appveyor executes from proj.4 main directory, not from proj.4/src)

(Windows builds use a different execution directory

The "return const err object" idiom (i.e.  const <type> err =
{HUGE_VAL,...}; ...  if (bad) return err) is problematic to implement
due to MSVC's misimplementation of HUGE_VAL as a non-const.

Hence, we need to run-time initialize these. In the pj_inv functions,
this was mistakenly done to the wrong object.

For pj_fwdobs/invobs and the remaining part of the obs-based API, this
is now worked around by providing functions returning a run time
HUGE_VAL initialized PJ_OBS or PJ_COO resp.

Obnoxious, but given MSVC's market penetration there is really not much
else we can do.

As per a review comment by Kristian Evers, @kbevers

Initializing after use in pj_fwd/pj_fwd3d. Forgotten #include horner.h
in PJ_horner.c