Evaluate observable rules in SI, not in declared units

[jeliason]

A rule is written in the model's declared units and is dimensionally consistent only after conversion. V_T adds V_Tmin in mL to a cell count times vol_cell in micrometre^3. SimBiology reconciles that by tracking units; the C++ emitter reconciles it by scaling every parameter on load (_class_parameter[P_vol_cell] = PFILE(QSP_vol_cell) * 602214.199). The Python emitter read the raw SBML numbers and applied nothing, so it was summing incompatible quantities.

Sharing an SBML parse was not enough to keep the two paths in step, because only one of them converted.

On the PDAC model all 13 constants in the live closure carry a factor, spanning 1.66e-24 to 6.02e5:

V_T at campaign p75:   2.16e13   ->  33.1 mL
stromal_fraction:      1.000000  ->  0.769  (and now varies with the patient)

That target could never move and would have read as total misfit against any real measurement.

Change: _classify reads get_si_factor for every parameter, species and compartment. emit converts inputs to SI, evaluates, and converts each requested output back, so the module takes and returns the model's own units. Factors appear as a visible SI_FACTOR table rather than folded into the arithmetic. Only inputs are required to declare units; an undeclared rule variable correctly defaults to 1.0.

Also stamps the emitter version plus a hash of its own source into the generated header, since the version alone does not move under an editable install.

Six new tests, including one pinning the exact failure (mL plus micrometre^3 giving 1.0 + 6e-12, not 7.0) and a native-units round trip.