[dmftproj] build fromfile basis representation in double precision

[krystophny]

Part of #7 (dmftproj Fortran-to-Python port).

The fromfile angular basis (a fromfile harmonics choice, e.g. a |j,m_j> basis) is read and represented in single precision, so case.symqmc/case.sympar carry a ~1e-7 error there. Two causes in set_ang_trans.f and setsym.f:

• CMPLX(rtrans, itrans) and EXP(CMPLX(0.d0, factor)) build complex values from REAL(KIND=8) arguments without a KIND, so the constructor evaluates in default (single) precision before widening to the COMPLEX(KIND=8) targets. Added KIND=8 to the basis-read and spinor-phase casts.
• the basis file is read through a 250-column buffer with an explicit buf1(2:250) slice, which truncates a higher-precision basis file mid-line. Widened the buffer and read the full line.

The surrounding code (the transmat/rotrep/rotl storage, the MATMULs, the read buffers) is already KIND=8; only the constructor intermediates leaked. The change makes them consistent.

No test regenerates from dmftproj (the converter tests read committed fixtures), so existing output is unaffected.

This is the single-precision path discussed on #148; it is a precision fix, not a fix for the symmetrization symptom reported there.

Verification

cmake + make + ctest (TRIQS, Build_Tests=ON):

100% tests passed, 0 tests failed out of 18

dmftproj on a full-precision |j,m_j> d-basis (16 operations, 8 time-reversal), against an independent double-precision generator:

before: max|fortran - ref| = 6.0e-8,  unitarity dev = 9e-9   (single precision)
after:  max|fortran - ref| = 1.9e-14, unitarity dev = 4e-16  (double precision)

Relates to #148.

[krystophny]

Superseded by #14, which extends this fromfile-only fix to every single-precision CMPLX cast in dmftproj (the spinrot phase in outputqmc.f, the Loewdin eigenvalue, the k-weights) plus full-precision cubic templates. With #14 the pure-Python port reproduces dmftproj to machine precision. Suggest closing this in favour of #14.

[krystophny]

Closing: superseded by #14, which fixes every single-precision CMPLX cast in dmftproj (not just the fromfile basis read) plus ships full-precision cubic templates. With #14 the pure-Python port reproduces dmftproj to machine precision.