[tmva][sofie] Improve AD-friendlieness of emitted code for Clad

[guitargeek]

This commit refactors SOFIE-generated inference code to enable correct and efficient reverse-mode automatic differentiation with Clad.

Key changes:

• Introduce explicit primitive operations (Copy, Fill, Relu) in SOFIE_common.hxx and provide corresponding custom pullbacks in CladDerivator.h. This replaces previously inlined loops and allows Clad to generate efficient gradient code without relying on tapes or loop-level differentiation.

• Update Gemm code generation to emit Copy/Fill instead of manually expanding bias initialization loops. This better exposes the intent and improves AD performance and correctness.

• Replace manual ReLU loops with a dedicated Relu() call, enabling a custom pullback that avoids tape-based condition tracking.

• Generate an additional "unoptimized" model variant in the SOFIE test suite (OptimizationLevel::kBasic), and use it for AD tests. This disables memory reuse of intermediate tensors. Opaque memory reuse is safe for inference but breaks source-transformation AD.

• Improve gradient test diagnostics in SOFIE Clad tests by reporting mismatched indices instead of only checking a global max difference.

With these changes, Clad-generated gradients for SOFIE models are both correct and significantly faster, reaching performance comparable to frameworks such as PyTorch and JAX on the CPU for the tested cases (fully-connected neural networks with multiple layers).

[github-actions]

Test Results

    22 files      22 suites   3d 7h 19m 23s ⏱️
 3 833 tests  3 831 ✅  1 💤 1 ❌
76 552 runs  76 533 ✅ 18 💤 1 ❌

For more details on these failures, see this check.

Results for commit c558240.

♻️ This comment has been updated with latest results.

[lmoneta]

LGTM!
Thank you, Jonas, for these improvements, which help CLAD. I have only a question on whether is better to impelment a Copy function or using directly std::copy.