new execution backend -> non-jit dynamic tensor execution

[martinjrobins]

It would be useful to have a non-jit based execution backend, one that just takes a DiscreteModel and executes it at run time by dynamically walking through the tensors and expressions, computing values as it goes.

This could execute diffsl models without any jit backend, just pure rust, so would be useful for pure wasm compilation.

This would also be useful to pre-calculate constant tensor values and inserting the values at compile time rather than using set_constants

[martinjrobins]

note: once this is done, remember to mark the constants array in llvm codegen as constant (currently non-constant), and remove the workaround introduced from CGMossa#3

[martinjrobins]

I think the most important part of this is to be able to pre-calculate the constants. Therefore I would propose to add a method TensorBlock::evaluate_constant that raises an error if the tensor block is not constant. If it is constant then it walks the expression tree evaluating the block and returning a Vec that is of length nnz, and containing the final calculated values of the tensorblock. This will need to handle all sorts of blocks, dense, sparse, diagonal, and also all sorts of contractions/broadcasting and Translations