Fix type inference

[frapac]

This PR allows to improve type inference in ExaTron. The signature of ExaTronProblem has been updated, from

mutable struct ExaTronProblem{VI, VD}

to

mutable struct ExaTronProblem{VI, VD, TM <: AbstractTronMatrix}

This allows Julia to infer directly the type of the matrix we are using, instead of dealing with unspecified AbstractTronMatrix.

Also:

• the preconditioner has been updated accordingly (we now specified the preconditioner directly in the option, :ICFS by default)
• we have fixed important type inference issue with TronDenseMatrix, leading to nice speed-up

A benchmark (using profiling/qptests.jl) gives:

• for TronSparseMatrixCSC, we reduce the overall number of allocations (now the function dtron is not allocating anything, which was not the case before, because of type inference issue)

#This PR
  369.747 μs (54 allocations: 2.36 KiB)
# main 
  386.941 μs (107 allocations: 3.59 KiB)

• for TronDenseMatrix{MT}, we get almost a 300x speed-up:

#This PR
  90.673 ms (69 allocations: 2.64 KiB)
# main 
  26.026 s (926692690 allocations: 13.81 GiB)

Fix #10

[frapac]

We need to specify explicitly the signature of the Array here, if we want to compile on Array{Float64, 2} instead of the more generic type Array

[youngdae]

Looks great! (Sorry for the delayed merging)