Release notes for TBCI NumLib 2.5.5
-----------------------------------
Version 2.5.5 has a number of fixes and enhancements over 2.5.4.

These are:
* Change the tarball to use .tar.bz2 instead of .tar.gz (tgz).
  Assign a version number there, also creat tarballs/zip files
  with autogenerated guiding decls / instantiation lists.
* TBCI::Matrix defined value_type to be aligned on a 16byte
  boundary. This could break (in svd_solver) on a 32bit
  machine. Took this as opportunity to clean up the typdefs:
  - value_type is now consistently just the base type of a 
    compound class (element_type aliasing value_type)
  - aligned_value_type may have additional alignment imposed
    for increased performance
* libtbcilapack did explicitly depend on libtbcidouble and
  implicitly as well on libtbcicplxdouble and libtbcistdcomplexdouble.
  These dependencies have been removed by adding a few conditional
  inline statements to BVector and F_TMatrix.
  libtbcilapack can now be used without the other tbci libraries.
  Note that if you previously relied on libtbcilapack pulling in 
  libtbcidouble at link time, you might need to change your Makefiles.
* In Matrix Vector multiplication, the TSMatrix variant had a bug
  where the result Vector would not have the right size for non-square
  matrices. (Likewise for F_TSMatrix.)
* memalloc_cache was not thread-safe; TBCI threads were thus not 
  supposed to do memory (de)allocations. The ones in TBCI didn't,
  but third parties may run into this trap.
  memalloc.h does now make sure only the main thread uses caching
  memory allocations. (This requires your platform to support thread
  local variables -- otherwise the use of threads will switch off
  memalloc_cache alltogether.) The downside of this is that for
  SMP=1 compiles, malloc_cache now has a link time dependency to
  smp.
  You can opt for using spinlocks instead by defining 
  TBCI_MALLOC_FORCESPINLOCK. I had expected this to yield better
  performance, but to my surprise this is not the case in general.
* Linux kernel 2.6.16 and newer report the total process CPU time
  when calling clock() from a thread (as they should following
  POSIX thread definition). Use clock_gettime() to get per thread
  CPU times if compiling with THREAD_STAT=1. This will need linking
  against librt.
* Minor fixes here and there (C_MEMALLOC, lapack's inv() routine ...)
* The specfile now does invoke a few regression tests to ensure the
  build's integrity.
* The special functions that have been converted with f2c a long time
  ago did not use prototypes consistently. This lead to a mismatch
  between integer and long types on 64bit platforms that went unnoticed
  by the compiler. Fixed that and made sure we use prototypes at least
  for the exported routines.
* The following fixes have been triggered by tests from Florian
  Hudelist, for which I'm very grateful.
* A new test program for calculating the eigenvalue in 
  lina/test/test_lapack_eig.cpp. It needs lapack, blas
  and the right fortran libs to be installed (libgfortran
  and libf2c or libg77 and libg2c).
* F_Matrix: This was implementation has obviously not seen much
  testing before. A few test programs exposed this and resulted
  in a code review that turned up a number of things. There was
  confusion of rows vs. coumns at a few places.