Allow current SciML dependency stack in tests

[ChrisRackauckas-Claude]

Please ignore this draft until it has been reviewed by @ChrisRackauckas.

Summary

• Allow ForwardDiff 1 in the root and AD test environments while retaining ForwardDiff 0.10.36 compatibility.
• Allow DataInterpolations 9 in the root test environment.
• Make the threaded ensemble regression test work with both SciMLBase v2's prob_func(prob, i, repeat) call and v3's documented prob_func(prob, ctx) call.
• Remove the stale downgrade comment from Re-enable downgrade (strict, allow_reresolve=false); intentionally RED pending upstream ForwardDiff/LogExpFunctions (deeper than the LinearSolve wall) — red until ForwardDiff 1.3.4 (#811) — natural failure, not disabled; auto-greens on upstream release. #307 now that the ForwardDiff/LogExpFunctions upstream blocker has shipped.
• Bump DiffEqCallbacks to 4.18.5.

Why

NonlinearSolve's DiffEqCallbacks downstream job failed before loading any changed NonlinearSolve code. Pkg.test() re-resolved DiffEqCallbacks' extras and found an empty intersection between developed NonlinearSolve's ForwardDiff-1-only requirement and DiffEqCallbacks' ForwardDiff-0.10-only test compat.

Fixing that first wall exposed two current-master compatibility issues behind it:

1. NonlinearSolve's SciMLBase 3.37 floor selects FindFirstFunctions 3, which requires DataInterpolations 9 rather than the test environment's previous 5–8 range.
2. SciMLBase 3 calls the public ensemble hook as prob_func(prob, ctx), while the existing regression test only accepted the SciMLBase 2 three-argument signature. The compatibility helper preserves distinct trajectory initial values using either the old integer index or public ctx.sim_id.

These are test-environment compatibility changes; package runtime dependencies are unchanged.

Investigation

• Inspected NonlinearSolve downstream job 90053534882.
• Reproduced its ForwardDiff resolver failure locally against clean NonlinearSolve upstream/master 1d51c8dac (v4.23.2), proving it is independent of PR #1100's code.
• Traced the ForwardDiff boundary to NonlinearSolve commit 9f60b2e29 (2025-12-27), which dropped root ForwardDiff 0.10 support; DiffEqCallbacks' 0.10-only test compat dates to c83fe3ed (2024).
• Replayed the downstream workflow after each wall, then fixed the newly visible DataInterpolations/FindFirstFunctions resolver boundary and SciMLBase v3 ensemble-hook test error.

Local validation

Julia 1.12.6:

• Exact downstream replay with clean NonlinearSolve master: passed (Testing DiffEqCallbacks tests passed).
• GROUP=Core: passed, including periodic tests 85/85 and saving tests 112/112.
• GROUP=AD: passed under ForwardDiff 1.4.2 (6/6, 16/16, and 1/1).
• GROUP=QA: passed (JET 15/15 and QA 21/21).
• Repository-wide Runic --check: exit 0.
• git diff --check: exit 0.

No test was skipped, silenced, weakened, or loosened.