* coefficient wise operators are more generic, with controllable result type.

- compatible with current STL's functors as well as with the extention proposal (TR1)
 * thanks to the above, Cast and ScalarMultiple have been removed
 * benchmark_suite is more flexible (compiler and matrix size)

doc/bench_unrolling

@@ -0,0 +1,11 @@
+#!/bin/bash
+
+# gcc : CXX="g++  -finline-limit=10000 -ftemplate-depth-2000 --param max-inline-recursive-depth=2000"
+# icc : CXX="icpc -fast -no-inline-max-size -fno-exceptions"
+
+for ((i=1; i<16; ++i)); do
+    echo "Matrix size: $i x $i :"
+    $CXX -O3 -I.. -DNDEBUG  benchmark.cpp -DMATSIZE=$i -DEIGEN_UNROLLING_LIMIT_OPEQUAL=1024 -DEIGEN_UNROLLING_LIMIT_PRODUCT=25 -o benchmark && time ./benchmark >/dev/null
+    $CXX -O3 -I.. -DNDEBUG -finline-limit=10000 benchmark.cpp -DMATSIZE=$i -DEIGEN_DONT_USE_UNROLLED_LOOPS=1 -o benchmark && time ./benchmark >/dev/null
+    echo " "
+done