fix(perf): retune Gemma1 decode

[inureyes]

Summary

Brings the older Gemma v1 hot path up to the same dtype-discipline + activation-shape level as the newer Gemma versions. On gemma-2b-4bit (M5 Max), decode throughput jumps from 69.34 → 214.49 tok/s — closing most of the gap to the mlx-lm baseline of 223.27 tok/s.

Cherry-picked from mlxcel-internal commit 1b4937a8. The internal benchmarks/, docs/model_tests_m5max.md, and docs_internal/performance/... paths from the original commit are intentionally excluded (none of them exist in this repo — same pattern as #20, #22).

What changed

src/models/gemma.rs — three coordinated fixes:

1. Quantized GeGLU activation now uses the same tanh-approx GELU path that mlx-lm picks for gelu_pytorch_tanh configs, instead of the exact-erf GELU. Bit-equivalent to the upstream reference.
2. Embedding scaling no longer promotes the whole activation tensor to fp32 via a scalar multiply — the scalar is cast to the activation dtype first, matching the pattern already used in gemma2 / gemma3.
3. Decode-time hints — enables the same layer-pipeline hint and maskless padded-prefill hooks the newer Gemma paths use, so Gemma v1 benefits from the same scheduler-level optimisations.

src/lib/mlxcel-core/src/utils.rs — 1-line supporting helper change to make the dtype-preserving scalar mul reusable.

Verification

• make verify-fmt — clean
• make verify-clippy (CI-faithful: --all-targets --features metal,accelerate -- -D warnings) — clean in 14s (warm cache)
• make verify-test skipped (15-30 min release run); the upstream commit is already validated against the M5 Max sweep with the throughput numbers above.