Reduce system to DAE index 0 in the presence of unbound inputs

[baggepinnen]

Following up on the discussion in https://github.com/SciML/ModelingToolkit.jl/issues/1303#issuecomment-960857224

It would be beneficial to have a version of dae_index_lowering/tearing that is aware of unbound_inputs in the sense of #1322
The motivation is detailed in the linked issue comment above, get a linear statespace system from a linearization function.

Most users would probably like to have this linearized system on the form

x' = Ax + Bu
y  = Cx + Du

but advanced users could use the index 1 form below as well

Ex' = Ax + Bu
 y  = Cx + Du

The problem with the current version of dae_index_lowering/tearing is that they do not distinguish between a state and a variable marked as [input = true] and thus fail to simplify the system.

[YingboMa]

To clarify, we only need to reduce away the linear algebraic equations in the linearized model instead of trying to solve nonlinear algebraic equations.

[baggepinnen]

This is now handled correctly by linearize