fastBPE step in embed.py fails if no tokenization

[bricksdont]

If there is no tokenization with embed.py, running the script fails:

cat input.de | python tools/laser/source/embed.py --encoder tools/laser/models/bilstm.93langs.2018-12-26.pt --bpe-codes tools/laser/models/93langs.fcodes --output embedded.de --verbose
 - Encoder: loading /net/cephfs/scratch/mathmu/laser-contra/tools/laser/models/bilstm.93langs.2018-12-26.pt
 - fast BPE: processing
Loading codes from /net/cephfs/scratch/mathmu/laser-contra/tools/laser/models/93langs.fvocab ...
fast: fastBPE/fastBPE.hpp:455: void fastBPE::readCodes(const char*, std::unordered_map<std::pair<std::__cxx11::basic_string<char>, std::__cxx11::basic_string<char> >, unsigned int, fastBPE::pair_hash>&, std::unordered_map<std::__cxx11::basic_string<char>, std::pair<std::__cxx11::basic_string<char>, std::__cxx11::basic_string<char> > >&): Assertion `splits.size() == 3' failed.
Aborted (core dumped)

This is due to the fact that the input file is sys.stdin before tokenization. The tokenizing step can handle input from STDIN, but the fastBPE step tries to execute the following command:

./fast applybpe [TWO SPACES HERE] /tmp/tmpgh9sgiy_/bpe \
    tools/laser/models/93langs.fcodes \
    tools/laser/models/93langs.fvocab

While the general recipe is

./fast applybpe output input codes vocab

The temporary output file /tmp/tmpgh9sgiy_/bpe is mistaken for the codes file, which makes this assertion fail.