Replace implementation of atomic addition.

Old version was based on atomicExch(), while this version uses CUDA's
built-in atomicAdd(), added in SM 2.0. When tested in isolation (test
code not provided in this commit), on a K10 (Kepler), the built-in
atomicAdd() is two times faster than the old version of atomic_add()
here, and on a 950M (Maxwell), 3 times faster.

Speed up to forward backward, however, is marginal for an
nnet3-chain-train call on the TEDLIUM version 1 dataset:

Times reported on a K10. Note speedup in BetaDashGeneralFrame(),
which is the only code calling the atomic add function.

New code:

[cudevice profile]
AddRows	0.468516s
AddVecVec	0.553152s
MulRowsVec	0.614542s
CuMatrix::SetZero	0.649105s
CopyRows	0.748831s
TraceMatMat	0.777907s
AddVecToRows	0.780592s
CuMatrix::Resize	0.850884s
AddMat	1.23867s
CuMatrixBase::CopyFromMat(from other CuMatrixBase)	2.04559s
AddDiagMatMat	2.18652s
AddMatVec	3.67839s
AlphaGeneralFrame	6.42574s
BetaDashGeneralFrame	8.69981s
AddMatMat	29.9714s
Total GPU time:	63.8273s (may involve some double-counting)
-----

Old code:

[cudevice profile]
AddRows	0.469031s
AddVecVec	0.553298s
MulRowsVec	0.615624s
CuMatrix::SetZero	0.658105s
CopyRows	0.750856s
AddVecToRows	0.782937s
TraceMatMat	0.786361s
CuMatrix::Resize	0.91639s
AddMat	1.23964s
CuMatrixBase::CopyFromMat(from other CuMatrixBase)	2.05253s
AddDiagMatMat	2.18863s
AddMatVec	3.68707s
AlphaGeneralFrame	6.42885s
BetaDashGeneralFrame	9.03617s
AddMatMat	29.9942s
Total GPU time:	64.3928s (may involve some double-counting)
-----

Updates to SRE08 example

nnet1: adding <ParametricRelu> component,

nnet3/report : Added plotting capability for parameter differences.

nnet1: minor cosmetic change,

WIP: Multi-database English LVCSR recipe

Removing little-used feature: time-reversed, and fwd-bkwd, decoding.

Various unrelated fixes: add --iter options to TIMIT sclite scoring; improve how syncfiles are removed in queue.pl; minor cosmetic and efficiency improvements in nnet3 code.

Extract wav - perturb_data_dir_speed.sh implementation

In scripts such as perturb-speed and perturb-volume scp lines are
tranformed into piped command with the appropropriate sox command. The
case that the scp file has file offsets was not handled. This commit
both generalizes the wav-copy command to work also on xfilenames and
fixes the two perturb scripts to use this command in case of file
offsets.

Speed up log softmax

New: For CuMatrix::LogSoftmax<float>, for dim = 16, speed was 0.0138019 gigaflops.
Old: For CuMatrix::LogSoftmax<float>, for dim = 16, speed was 0.0133804 gigaflops.
New: For CuMatrix::LogSoftmax<float>, for dim = 32, speed was 0.056202 gigaflops.
Old: For CuMatrix::LogSoftmax<float>, for dim = 32, speed was 0.052121 gigaflops.
New: For CuMatrix::LogSoftmax<float>, for dim = 64, speed was 0.227829 gigaflops.
Old: For CuMatrix::LogSoftmax<float>, for dim = 64, speed was 0.186255 gigaflops.
New: For CuMatrix::LogSoftmax<float>, for dim = 128, speed was 0.65638 gigaflops.
Old: For CuMatrix::LogSoftmax<float>, for dim = 128, speed was 0.65072 gigaflops.
New: For CuMatrix::LogSoftmax<float>, for dim = 256, speed was 2.15268 gigaflops.
Old: For CuMatrix::LogSoftmax<float>, for dim = 256, speed was 1.64888 gigaflops.
New: For CuMatrix::LogSoftmax<float>, for dim = 512, speed was 5.1179 gigaflops.
Old: For CuMatrix::LogSoftmax<float>, for dim = 512, speed was 3.85136 gigaflops.
New: For CuMatrix::LogSoftmax<float>, for dim = 1024, speed was 10.8209 gigaflops.
Old: For CuMatrix::LogSoftmax<float>, for dim = 1024, speed was 6.76963 gigaflops.
New: For CuMatrix::LogSoftmax<double>, for dim = 16, speed was 0.0133584 gigaflops.
Old: For CuMatrix::LogSoftmax<double>, for dim = 16, speed was 0.011373 gigaflops.
New: For CuMatrix::LogSoftmax<double>, for dim = 32, speed was 0.0533796 gigaflops.
Old: For CuMatrix::LogSoftmax<double>, for dim = 32, speed was 0.0528196 gigaflops.
New: For CuMatrix::LogSoftmax<double>, for dim = 64, speed was 0.202721 gigaflops.
Old: For CuMatrix::LogSoftmax<double>, for dim = 64, speed was 0.170107 gigaflops.
New: For CuMatrix::LogSoftmax<double>, for dim = 128, speed was 0.627234 gigaflops.
Old: For CuMatrix::LogSoftmax<double>, for dim = 128, speed was 0.722198 gigaflops.
New: For CuMatrix::LogSoftmax<double>, for dim = 256, speed was 1.89987 gigaflops.
Old: For CuMatrix::LogSoftmax<double>, for dim = 256, speed was 1.44478 gigaflops.
New: For CuMatrix::LogSoftmax<double>, for dim = 512, speed was 4.14807 gigaflops.
Old: For CuMatrix::LogSoftmax<double>, for dim = 512, speed was 3.37973 gigaflops.
New: For CuMatrix::LogSoftmax<double>, for dim = 1024, speed was 6.70849 gigaflops.
Old: For CuMatrix::LogSoftmax<double>, for dim = 1024, speed was 4.96657 gigaflops.

nnet3: lstm/make_configs.py : Removed a bug where label_delay was not…

nnet3: lstm/make_configs.py : Removed a bug where label_delay was not being added to the xentropy branch in chain models.

add frame_shift_opts to score

Fix for online-decoder crash reported in PR #993 by @fanskyer

Add unit test for CuMatrix::AddMatMatDivMat

Fix oracle calculation for nnet3 scripts

-removed --per-utt from nnet3/decode.sh
-removed lstm/decode.sh from local/nnet3/run_lstm.sh
-reverted adding oracle.sh to nnet3

Exit on incorrect num arguments in validate_data_dir.sh