fix(admin/benchmark): forward SpecPrefill model settings to engine

[SuperMarioYL]

Summary

Fixes #1145.

The dashboard benchmark path read ModelSettings.specprefill_enabled only
to tag run.experimental_features, but never forwarded the matching
specprefill_keep_pct / specprefill_threshold overrides to
engine.stream_generate (in _run_single_test) or to
engine_core.add_request (in _run_batch_test).

The chat-completion path already loads these from ModelSettingsManager
(see server.py around the per-request override block),
so end-users see SpecPrefill engage at their configured threshold during
normal chat — but a benchmark for the same model silently falls back to
DEFAULT_THRESHOLD=8192 and reports numbers that don't reflect the
configured threshold (e.g. the issue's reproducer: threshold set to 512,
benchmark only engages SpecPrefill above 8192).

What changed

omlx/admin/benchmark.py:

• Build a specprefill_kwargs dict once at run-start, alongside the
  existing experimental-features snapshot.
• Add an optional specprefill_kwargs param to _run_single_test and
  _run_batch_test and forward via **kwargs into
  engine.stream_generate(...) / engine_core.add_request(...).
• Pass specprefill_kwargs through to the warmup call too, so the
  warmup path doesn't load a different code branch than the measured
  runs.

When the model has no SpecPrefill settings the dict stays empty and no
specprefill_* kwarg leaks into stream_generate — preserving the
default engine path for models the user has not opted in for.

tests/test_benchmark.py:

• New TestSpecPrefillSettingsLoad class with three regression tests
  that mock stream_generate / add_request and assert the threshold
  and keep_pct round-trip end-to-end through the benchmark helpers, plus
  a no-leakage case for models without SpecPrefill enabled.

Test plan

• python -m py_compile omlx/admin/benchmark.py tests/test_benchmark.py
• Isolated test run of the four TestSpecPrefillSettingsLoad cases
  (stubbed mlx.core to import omlx.admin.benchmark standalone on
  a Linux dev machine without the MLX runtime) — all pass on the
  branch and the assertions fail on main as expected.
• Full pytest tests/test_benchmark.py in the project MLX
  environment (not runnable in this dev environment without the
  Apple Silicon MLX runtime).

Reviewer kill-question (anticipated)

"Why not push the settings-manager load into engine.stream_generate
itself?"

The engine layer is intentionally settings-manager-agnostic — every
caller (chat-completion, benchmark, tests) forwards the per-request
override. Pushing settings-manager into the engine would duplicate the
existing forward-flow and add a new dependency cycle from
omlx/engine/* into omlx.model_settings. The benchmark route is the
only layer that was missing the forward step, so the fix lives there.