High-precision accuracy measurements

The scripts in this directory measure the relative error of libcerf
functions against ball-arithmetic references (python-flint / Arb). The
figures quoted in the ACCURACY sections of the man pages come from them:

  wofz_relerr.py             w_of_z at the generated test points and at random
                             points beyond |z| = 7
  cheb_relerr_testpoints.py  erfcx, im_w_of_x, voigt_hwhm at their generated
                             test points, split by code path
  cheb_relerr_random.py      the same at random points over wide ranges
  voigt_hwhm_large_ratio.py  voigt_hwhm for gamma/sigma between 1e3 and 1e8
  wofz_components.py         re_w_of_z and im_w_of_z close to the two axes,
                             where the modulus criterion does not bind (#29)
  wofz_near_axis.py          the expansion of Re w about the real axis used by
                             w_of_z: recurrence, order, and its limit (#29, #30)
  expz2_relerr.py            the six functions carrying the factor exp(+-z^2),
                             whose exponent is a double-word (#32)
  wofz_zeros.py              the zeros of w below the real axis, how far their
                             asymptotic expansion has to be carried, the error
                             of w_of_z close to one of them, and what a phase
                             form would achieve instead; also generates the
                             near-zero test table
  wofz_phase.py              that phase form prototyped in double-word
                             arithmetic, operation for operation as the
                             library would compute it (#36)
  wofz_anchors.py            the anchored polynomials that serve the ten
                             innermost zeros, and the table they live in

Requirements: python3 with python-flint; the library under test, found
through LIBCERF_PATH or the system library search; the test tables in
../../test and the reference modules in ../ppapp. Each script prints the
maxima it finds; its docstring says which man-page figures depend on it.

One table is kept here rather than recomputed: wofz_zeros.tab, the first
500 zeros of w in quadruple precision, written by

    python3 wofz_zeros.py table 500 6 > wofz_zeros.tab

and so is the test table of the grid points around the zeros:

    python3 wofz_zeros.py tests > ../../test/auto_test_zeros_wofz.c

and so is the table of anchored polynomials:

    python3 wofz_anchors.py table > ../../lib/auto_anchor_wofz.c

The broader statistics of the w_of_z page (10^9 random points inside
|z| < 7, 3e9 beyond) come from ttapp modes e and H; see the cgt
manuscript and its data directory.
