- supports arbitrary number of languages,
- uses 'multi-task' training with per-language weights (one softmax per language),
- (the decoding script not yet merged)

- in case a task gets 0 frames, use 0.0 as objective value, avoid nan,

Addressing @danpovey comments: svn is still a prerequisite; svn mirror is not maintained; version control paragraph in tutorial_code.dox rewritten.

Documentation changes: added Git tutorial, removed Subversion tutorial and updated multiple references from Subversion to Git and from SourceForge to Kaldi's own web site or GitHub as appropriate.

Modifying doc/README with method of uploading to sourceforge that works when their shell access is down.

Setting the windows/generate_solution.pl parameters to be in sync in documentation and with the *props files

windows fixes -- resolving issues with incrrectly behaving expf/exp(float) under MSVC2013, plus some other improvements

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@5243 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

trunk: modifying cu-device.cc to work around what I'm not sure is a bug or a feature in our installation of CUDA toolkit v7.0.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@5241 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8