Preserve identity-distinct autograd views in StreamingAttentionKLDivergenceReducer and add SSHR integration test

[stephandooper]

Motivation

• Ensure streaming KL reducer presents identity-distinct views of its inputs (matching AttentionKLDivergenceReducer passthrough) so reducer head resolution can disambiguate repeated tensor occurrences in flattened outputs.
• Preserve the original autograd graph for downstream backward replay by avoiding cloning or detaching the tensors.
• Add a focused integration regression that reproduces SSHR's layout where attention tensors are reused as auxiliary inputs and later as primary inputs to KL reducers.

Description

• Modify StreamingAttentionKLDivergenceReducer.forward to return identity-distinct views using view_as, store the pair in _last_inputs, set _last_output to the student view, and return the pair; no cloning or detaching is performed. (file updated: lightstream/core/reducer/kl.py).
• Add a small SSHR-like test fixture SSHRReducerLayoutNet and a regression test_sshr_layout_maps_kl_reducers_to_primary_occurrences that builds the flattened outputs and asserts KL reducer head indices and output shapes. (file updated: tests/test_scnn.py).
• The change keeps autograd connectivity intact while giving each view a unique Python identity so _resolve_reducer_head_map can map reducers to their primary occurrences instead of earlier auxiliary occurrences.

Testing

• Ran a focused Python check that instantiated StreamingAttentionKLDivergenceReducer, asserted the returned views are distinct objects, verified _last_inputs/_last_output, and confirmed gradients propagated through the original tensors; this check passed.
• Ran bytecode/format checks via python -m compileall -q lightstream tests/test_scnn.py which succeeded.
• Ran the focused pytest tests/test_scnn.py::test_sshr_layout_maps_kl_reducers_to_primary_occurrences which exercised the new integration test, but it failed in this environment due to absence of an NVIDIA driver (CUDA initialization error) despite attempting to run with CPU/statistics settings; the failure is environmental rather than indicating a logic regression.

---

Codex Task