Ported fixes: iEMG SFAP units and execution route, fibre CV units, Gfluctdv noise - #43

… fibre layouts

Fibre CV used a slope of 5.5 (m/s)/mm with diameters in mm, giving ~1.1 m/s
before the 2 m/s clip, so every fibre had exactly 2.0 m/s. Use CV = 0.95 m/s
+ 0.05 m/s per um (Nandedkar & Stalberg 1983), shared with
SurfaceFiberDispersion through CV_INTERCEPT__M_S / CV_PER_DIAMETER__M_S_PER_MM.
Affects intramuscular EMG; surface EMG uses the mean CV.

voronoi_pi1e5.csv covers a square, so only ~313 k points form a full disc.
Larger fibre counts got a corner-only outer ring and, above 399 499, a silent
cap. Those counts now use a scrambled Halton sequence in the disc (seeded from
MyoGen's RNG) with a warning; smaller counts are unchanged.

Fresh-trunk fixed both differently (be9cdfe35, a82923b2b); see
docs/pr/2026-11-01-upstream-changes.md for why its numbers were not used.

…ation

Four errors in the intramuscular source and volume conductor nearly cancelled
at the default 0.01 mm spatial step:

- muscle conductivities defaulted to 63 / 330 "S/m" (1000x tissue values);
  now 0.063 / 0.33 S/m, converted to S/mm (bioelectric.py, fiber_simulation.py)
- the Rosenfalck IAP was evaluated at 2x the distance with the chain rule
  applied twice (half-length IAP, 8x current); now lambda = 1 mm, as in sEMG
- the wave towards the shorter fibre half was subtracted, inverting it; the
  two waves now add (a second derivative is unchanged by reflection)
- calc_sfaps summed over z without the spatial step, so amplitude scaled with
  1/dz; both CPU and GPU paths now multiply by dz

A single-fibre SFAP through MotorUnitSim.calc_sfaps now matches an independent
SI line source within 3 % at 0.1-1 mm, is mirror-symmetric about the
end-plate, independent of dz, and identical on GPU and CPU
(tests/test_iemg_sfap_units.py). Output documented in mV.

runtime/density.py is a numba port of bioelectric.get_current_density
and still had the pre-fix kernel: IAP at twice the distance with the
chain rule applied twice, the shorter-half wave subtracted, and
runtime/iemg.py summed the line source without dz. MUAPs from
execution=ExecutionPolicy() disagreed with the default route (peak ratio
~1/8, correlation ~0.09 on a single fibre).

Pin the kernel to the reference with an equivalence test, add a
default-route vs execution-route MUAP test, and bump
IEMG_CACHE_NAMESPACE to v3 so stale cached templates are recomputed.
CHANGELOG entries for the ported fixes and the SFAP units.