Support untied-head attention-sink LlamaSimpleMLP targets

[claude-spd1]

Description

Add the untied-output-head and learned attention-sink variant of the 4-layer LlamaSimpleMLP target to the public loader and pretrainer. Clean and decomposed forwards share the same sink-aware attention implementation, and the cache contract records both new model facts.

Related Issue

Bridge task #1423.

Motivation and Context

The completed #1423 decomposition and base-model bundles need a public consumer that reconstructs their exact architecture. Without this change, public param-decomp assumes a tied output head and plain causal attention and cannot load those artifacts faithfully.

How Has This Been Tested?

• pytest -q param_decomp/tests/core/test_pretrain.py param_decomp/tests/targets/test_llama_simple_mlp.py (21 passed, 1 skipped)
• pytest -q param_decomp/tests/core/test_config.py param_decomp/tests/targets/simple_mlp_equivalence/test_torch_equivalence.py (22 passed, 1 skipped)
• The cache round-trip test now covers both the legacy tied/plain variant and the untied/sink variant.

Does this PR introduce a breaking change?

No. Existing configs default to tied embeddings and no attention sinks.

Diff breakdown

• pretrain: +98 -25
• target loader and attention primitive: +151 -20
• tests: +24 -2
• docs: +3 -2

Crew-Address: agent/i9y2
Bridge-Task: 1423