Compute TSRKC2/TSRKC3 stage time unconditionally in the stage loop

Rebase of the surviving content of #3902 onto current master.

In the in-place TSRKC2 and TSRKC3 stage loops `th` was assigned inside
the `if (iter < mdeg)` guard. Hoisting it out matches RKC and RKMC2,
which already compute `th` unconditionally.

This is a no-op: `th` is read only by the `th1 = th` on the following
line, which stays inside the guard, so nothing observes the value
computed on the final iteration.

The rest of #3902 landed separately and is already on master:
  - RKL1/RKL2 non-autonomous stage times (#3948)
  - RKG1/RKG2 non-autonomous stage times (#3949)
  - RKC in-place mdeg < 2 handling (#3951)
  - per-algorithm limiter fields, reverted by the author in c34fe7a
    and superseded by solve-level limiters (#3835)

Original commits: 9158496, 05548a5, c34fe7a

Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>

Author: Sreeram-Shankar

lib/OrdinaryDiffEqStabilizedRK/src/rkc_perform_step.jl

@@ -1226,10 +1226,10 @@ end
         f(k, gprev, p, t + dt * th1)
         OrdinaryDiffEqCore.increment_nf!(integrator.stats, 1)
         @.. broadcast = false u = μ * gprev + ν * tmp + dt * μs * k
+        th = μ * th1 + ν * th2 + μs
         if (iter < mdeg)
             @.. broadcast = false tmp = gprev
             @.. broadcast = false gprev = u
-            th = μ * th1 + ν * th2 + μs
             th2 = th1
             th1 = th
             z2 = z1
@@ -1486,10 +1486,10 @@ end
         OrdinaryDiffEqCore.increment_nf!(integrator.stats, 1)
         @.. broadcast = false u = μ * gprev + ν * tmp + (T(1) - μ - ν) * uprev +
             dt * μs * (k - νs * fsalfirst)
+        th = μ * th1 + ν * th2 + μs * (T(1) - νs)
         if (iter < mdeg)
             @.. broadcast = false tmp = gprev
             @.. broadcast = false gprev = u
-            th = μ * th1 + ν * th2 + μs * (T(1) - νs)
             th2 = th1
             th1 = th
             b2 = b1