fix(py): make subthreshold_swing fail loudly, and state its units

[reneotten]

fix(py): make subthreshold_swing fail loudly, and state its units

Two problems, both about the same function.

1. Silent NaN. Device() defaults to v_ds = 0, where the Landauer
   integrand vanishes and every swept current is exactly zero. The
   log10(I) fit then produced -inf, and 1/-inf returned nan with only
   a numpy RuntimeWarning. Device().sweep_v_g(0, 0.5, 0.05).subthreshold_swing()
   quietly handed back nan instead of explaining that a drain bias is
   needed. Non-positive currents, empty fit windows and degenerate slopes now
   raise ValueError naming the likely cause.

2. Contradictory units. sweep.rs described the result as "mV/decade
   equivalent, in the original's V/decade units", which asserts two different
   units in one clause, and the Python docstring gave none at all. The value
   is volts/decade; the only place that was resolved was a * 1000 in a
   notebook cell.

Both doc comments now say V/decade and give the 300 K thermal limit
(0.0596 V/decade) as a sanity anchor. Verified against the model: the
default device gives 65.8 mV/decade at 300 K and 43.8 mV/decade at 200 K,
each just above the corresponding ln(10)kT/e limit, which is the expected
behaviour for a thermally-limited barrier.