[WIP] Laser forEach iterators

[mratsim]

This replaces the use of mapX_inline and applyX_inline by the forEach / forEachContiguous / forEachParallel / forEachSerial laser iterators.

This is particularly valuable for recurrent neural network like GRU because we can implement the equation in a straightforward manner with meaning ful variable name instead of magic x, y, z and we would have needed an apply11_inline anyway (with the correct var/non-var parameters).

TODO:

• There is a significant parallel performance regression on GRU when running test_nnp_gru. While before this PR the parallel version was "only" 2x slower than serial, now it's 13x slower than serial which probably signals a false sharing issue.
  Even then RNN are a type of iterative stencil application that require special care and often used for polyhedral benchmarking so another approach using tiling is probably needed to properly speed those up. (see https://github.com/numforge/laser/blob/master/research/automatic_loop_nest_scheduling.md#polyhedral-approaches)

Not in scope:

• using forEach for backward propagation in GRU: this is a headache inducing refactoring

[mratsim]

Some changes in operator BLAS L1, require changing autograd to +.= because the += based on apply2_inline was doing implicit broadcast somehow. But then we have cascading issues that requires changing +.= to use broadcast2 and broadcast2 fixes, and then we have the Stack autograd layer that is failing tests. See 2929cb5 in the laser-iterators-stashed branch