(grafted from c60b58edddc954a2a712373393a669ac7ef349cf)

Refined the #ifdef __CUDACC__ guard to ensure that when trying to compile gpu code with a non cuda compiler results in a linking error instead of bogus code.

Bug 1089: add a warning when using a MatrixBase method which is implemented within another module by declaring them inline.

Bug 1090: fix a shortcoming in redux logic for which slice-vectorization plus unrolling might happen.

Improve numerical accuracy in LLT and triangular solve by using true scalar divisions (instead of x * (1/y))

Updated the custom indexing code: we can now use any container that provides the [] operator to index a tensor. Added unit tests to validate the use of std::map and a few more types as valid custom index containers

Tightened the definition of isOfNormalIndex to take into account integer types in addition to arrays of indices
Only compile the custom index code  when EIGEN_HAS_SFINAE is defined. For the time beeing, EIGEN_HAS_SFINAE is a synonym for EIGEN_HAS_VARIADIC_TEMPLATES, but this might evolve in the future.
Moved some code around.

and base nested_eval on it.

Update custom setFromTripplets API to allow passing a functor object, and add a collapseDuplicates method to cleanup the API. Also add respective unit test