[inureyes]

Summary

The single-sequence XLA/IREE decode graph contracted each projection's last axis against the checkpoint's [out, in] weight layout. IREE lowered that to a full-weight transpose plus a vecmat on every decode step: profiling on gfx1151 showed 112 projection transposes (57 ms) and a 9.65 ms f32 LM-head transpose inside a 96.5 ms step, against 4.1 ms/step for the MLX ROCm path (issue #2859).

This PR makes the graph declare projection weights [in, out] and contract the shared axis, so IREE emits one vecmat per projection, and moves the transpose to load time where it runs once per weight.

Changes

• ProjectionLayout (OutIn | InOut) on Config and Builder, with linear_proj / linear_seq_proj selecting the contraction to match the declared weight shape. Dense projection call sites moved to the layout-aware helpers; Phi4MM LoRA adapter tensors (modal, [rank, in] / [out, rank]) and the LM head (Whole) keep the untransposed helpers.
• transpose_projection in weights.rs, a 32x32 blocked [out, in] -> [in, out] transpose applied once per Proj / Rows weight at upload. The packed 4-bit path keeps its in-graph dequant plus transpose, since it already trades decode throughput for upload bytes.
• The layout is chosen per engine, not globally: IreeLlama (single sequence) emits InOut; IreeRaggedLlama (continuous batching) keeps OutIn; Gemma3n's own emitters keep OutIn.

Why the layout is per engine

The best lowering depends on the graph shape, not the model. The ragged batch decode already lowered the old [out, in] contraction to a single transposed-B linalg.matmul (about 0.3 ms per down projection at its row count), while the canonical matmul over [in, out] was about 8x slower per call (2.38 ms); a single-layout branch measured B=4 N=8 at 28.0-28.5 tok/s against 59.3-65.5 before. The single-row case is the opposite, where the old layout cost the transpose plus vecmat. Keeping OutIn for the batching engine also keeps every committed golden byte-identical, which is what the 385-test mlxcel-xla lib suite checks.

Validation (gfx1151, hip, llama-3.2-1b-instruct-4bit, release build)

case	before	after
B=1 CLI, n=64, capacity 256	10.5-11.4 tok/s	18.6-21.4 tok/s
B=1 CLI, n=128, capacity 256	9.0-10.0 tok/s	16.5-18.3 tok/s
batch B=4 N=8	63.9-69.2 tok/s	66.8-67.7 tok/s
batch B=8 N=16	80.1-80.5 tok/s	74.8-84.1 tok/s

• REFERENCE-EXACT PASS on both batch sizes.
• Greedy output token-identical to the previous binary on llama-3.2-1b, qwen3-1.7b, qwen2.5-0.5b and gemma-3-4b.
• cargo test -p mlxcel-xla --lib 385 passed, committed goldens unchanged.
• cargo clippy -p mlxcel --lib --tests -- -D warnings and cargo clippy -p mlxcel-xla --all-targets --features iree -- -D warnings clean; cargo fmt --check clean.
• cargo check --features xla-iree --bin mlxcel --example xla_batch_bench clean.

Closes #2859