Make forwardDims() optional and handle data dependency

Should address and close #16, #48 (closed), #92 and #95:

• #16: re-focus the goal of compile() function, less centered around forwardDims().
• #48 (closed): remove the associateInput() execution path of forwardDims(), therefore decoupling the two mechanisms, leaving much clearer error checking.
• #92: forwardDims() does not fail anymore if some operator cannot compute output dims (e.g. GenericOperator), but simply return false.
• #95: enable scheduling of output dims data dependent graphs using tokens, by making forwardDims() optional in this case.

Changes:

• Renamed computeOutputDims() to forwardDims(), for consistency with GraphView::forwardDims().
• Added allowDataDependency to forwardDims(), for handling of output dims data dependent operators.
• Removed associateInput() from forwardDims(): only keep assertions, as input association should be properly handled outside this function.
  • Adapted Identity operator to work properly without forwardDims().
• Added forwardDims() call in OperatorTensor::forward() if output dims are not forwarded.
• forwardDims() is now fully optional, removed the call when not necessary.
• Added default implementation for several operators, see &2.