fix(speed_benchmark): align benchmarks with the Reactive_E2E contract (#92)

[gcordova10]

Problem

Model/speed_benchmark/speed_benchmark.py and planner_benchmark.py still assume
the pre-refactor API and crash on the current main (post Reactive_E2E refactor).

Fix

• speed_benchmark.py
  • AutoE2E no longer accepts fusion_mode (concat / cross_attn and the knob were
    removed; BEV is the only multi-view fusion) → construct with
    view_fusion_kwargs={"bev_h": 8, "bev_w": 8} and drop the fusion_mode sweep.
  • AutoE2E.forward now takes map_input and returns the trajectory tensor
    directly (the (loss/traj, ego_hidden, future) 3-tuple was removed) → feed a
    map image and stop unpacking a tuple.
• planner_benchmark.py
  • gru was removed from PLANNER_REGISTRY → benchmark flow_matching vs bezier.
  • planners now return the trajectory directly, so read out (not out[0]) for the
    smoothness metric.

Verification

• planner_benchmark.py runs end-to-end: bezier jerk ~5e-6 vs flow_matching ~3.4
  (the Bernstein-basis smoothness guarantee holds, as expected).
• speed_benchmark.py forward path validated on CPU (is_pretrained=False).
• ruff + mypy clean.

Closes #92.

[gcordova10]

CI green. The two harnesses under Model/speed_benchmark/ were broken against current main after the Reactive_E2E refactor; this aligns them with the live contract:

• speed_benchmark.py: AutoE2E no longer accepts fusion_mode (concat/cross_attn and the knob were removed; BEV is the only multi-view fusion) → construct with view_fusion_kwargs={bev_h, bev_w}. forward now takes map_input and returns the trajectory tensor directly (the (loss/traj, ego_hidden, future) 3-tuple was removed) → feed a map image, stop unpacking.
• planner_benchmark.py: gru was removed from PLANNER_REGISTRY → benchmark flow_matching vs bezier; planners return the trajectory directly, so read out (not out[0]) for the smoothness metric.

Verified end-to-end (planner_benchmark: bezier jerk ~5e-6 vs flow_matching ~3.4 — the Bernstein-basis smoothness guarantee holds); speed_benchmark forward path validated on CPU. ruff + mypy clean. Closes #92.