Improve root-finding convergence with bisection and log-scale formulation

[aarmey]

Summary of Changes

• 1D Bracketed Bisection for polyfc: Replaced unconstrained Levenberg-Marquardt with optimistix.Bisection on the bounded domain [0, Phi_max]. Because the residual function is strictly monotonic and bounded, this provides guaranteed 100% convergence with no overshooting or domain errors.
• Log-Space Parameterization for polyc: Updated Req_polyc to solve for log(Req) using the log-ratio residual log(Req + Rbound) - log(Rtot). This prevents negative receptor concentrations during intermediate iterations and stabilizes gradients across wide dynamic ranges.
• Analytical 1:1 Langmuir Initial Estimate for Req_solve: Initialized the solver with the 1:1 binding equilibrium estimate instead of a zero vector, starting the solver near the true root and reducing iterations.
• New Edge Case Tests: Added regression tests in valentbind/test/test_edge_cases.py covering high-affinity/high-valency multivalent mixtures (f=12, wide affinity ranges).

All 16 unit tests, static type checks (make ty), and Ruff lint/formatting checks pass.