Revert SUNDIALS curvature threshold, use formula-always approach

The SUNDIALS yddnrm*hub^2>2 threshold is designed for sqrt(2/yddnrm)
(p+1=2) but doesn't work for order-dependent (2/yddnrm)^(1/(p+1)).
High-order methods (p=5+) almost never trigger the formula, falling
back to tiny geometric mean steps.

Use formula-always-with-hub-clamp: when yddnrm > 0 compute the
order-dependent step and clamp to hub; when yddnrm == 0 use hub.

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: ChrisRackauckas

lib/OrdinaryDiffEqCore/src/initdt.jl

@@ -395,10 +395,6 @@
             yddnrm = internalnorm(tmp, t) / hg * oneunit_tType
 
             # Order-dependent step proposal: h ~ (2/yddnrm)^(1/(p+1))
-            # Always use the formula and clamp to hub, rather than falling back
-            # to the conservative geometric mean sqrt(hg*hub). This prevents
-            # overly small initdt for high-order explicit methods where the
-            # formula naturally gives hnew > hub.
             if DiffEqBase.value(yddnrm) > 0
                 hnew = convert(
                     _tType,