Having some basic BLAS operations implemented in MLlib can help simplify the current implementation and improve some performance.

Tested on my local machine:

~~~
bin/spark-submit --class org.apache.spark.examples.mllib.BinaryClassification \
examples/target/scala-*/spark-examples-*.jar --algorithm LR --regType L2 \
--regParam 1.0 --numIterations 1000 ~/share/data/rcv1.binary/rcv1_train.binary
~~~

1. before: ~1m
2. after: ~30s

CC: jkbradley

Author: Xiangrui Meng <meng@databricks.com>

Closes #1849 from mengxr/ml-blas and squashes the following commits:

ba583a2 [Xiangrui Meng] exclude Vector.copy
a4d7d2f [Xiangrui Meng] Merge branch 'master' into ml-blas
6edeab9 [Xiangrui Meng] address comments
940bdeb [Xiangrui Meng] rename MLlibBLAS to BLAS
c2a38bc [Xiangrui Meng] enhance dot tests
4cfaac4 [Xiangrui Meng] add apache header
48d01d2 [Xiangrui Meng] add tests for zeros and copy
3b882b1 [Xiangrui Meng] use blas.scal in gradient
735eb23 [Xiangrui Meng] remove d from BLAS routines
d2d7d3c [Xiangrui Meng] update gradient and lbfgs
7f78186 [Xiangrui Meng] add zeros to Vectors; add dscal and dcopy to BLAS
14e6645 [Xiangrui Meng] add ddot
cbb8273 [Xiangrui Meng] add daxpy test
07db0bb [Xiangrui Meng] Merge branch 'master' into ml-blas
e8c326d [Xiangrui Meng] axpy