feat(xla): add prefill embeddings entry (#858)

[inureyes]

Summary

Adds a distinct prefill_embeddings.main StableHLO entry that accepts post-scale hidden states and an explicit additive attention bias while preserving the existing token-prefill output and KV-cache contract.

Changes

• Refactors token and embeddings prefill through one shared transformer, normalization, LM-head, output-selection, and KV path.
• Defines a stable embeddings, positions, real_len, attention_bias schema with fail-closed metadata and finite-mask validation.
• Preserves tied and untied LM heads, freezes an untied embeddings-prefill golden, and keeps all existing token-prefill goldens byte-identical.
• Adds a real IREE llvm-cpu parity harness covering tied and untied heads, nonzero padding, real_len = 1, and near-capacity input.

Validation

• cargo test -p mlxcel-xla --lib prefill_embeddings -- --nocapture (3 passed, 1 opt-in dump test ignored)
• cargo test -p mlxcel-xla --lib validation::tests::registered_fixtures_are_byte_exact -- --nocapture
• scripts/xla/validate_arch.sh --structural-only (9 architecture checks passed)
• cargo check -p mlxcel-xla --lib --tests
• cargo clippy -p mlxcel-xla --lib --tests -- -D warnings
• cargo fmt --all -- --check
• spike/openxla/.venv/bin/python spike/openxla/prefill_embeddings_check.py (exact zero difference for logits and full K/V outputs in every scenario)

Review

Correctness review confirms deterministic argument order, explicit mask polarity and padding semantics, tied/untied head selection, scalar argmax and logits variants, and byte-exact preservation of the token entry. Security review confirms malformed shape, dtype, length, NaN, infinity, and ambiguous mask values fail closed. Performance review found no new copies on the embeddings hot path; validation performs one bounded scan of the static bias before compilation or invocation.

Closes #858