feat: Device eval branch specialization (Phase 1)

[robtaylor]

Summary

• Inline all shared parameter values (2,591 for PSP103) and shared cache values (407) as Python float literals in the generated eval function code
• At JAX trace time, jnp.where(const_bool, a, b) only traces the taken branch, eliminating static-parameter-dependent branches from the compiled XLA program
• Add NR phase timing profiler (scripts/profile_nr_phases.py) and parallelism analysis tooling (scripts/analyze_parallelism.py)

How it works

Before (traced): XLA sees abstract values, evaluates both jnp.where branches:

v123 = shared_params[42]        # AbstractValue
v456 = v123 > 0.0               # AbstractValue (bool)
result = jnp.where(v456, a, b)  # Both a and b traced

After (concrete): Python evaluates constants, only one branch traced:

v123 = 1.5e-6                   # Python float literal
v456 = v123 > 0.0               # Python True
result = jnp.where(v456, a, b)  # Constant-folds to just a

Verification

Generated PSP103 eval code shows 0 shared_params[N] references and 0 shared_cache[N] references (all 2,998 values inlined as literals).

Benchmarks produce identical numerical results:

Benchmark	Model	Result
rc	resistor+capacitor	Identical
graetz	diode	Identical
ring	PSP103 (NMOS+PMOS)	Identical

Test plan

• compare_vacask.py --benchmark rc passes
• compare_vacask.py --benchmark graetz passes
• compare_vacask.py --benchmark ring passes
• GPU profiling on CUDA (c6288, mul64) for performance delta