Fix set_timesteps: torch.flip instead of tensor [::-1]

[alex-16moro]

Summary

set_timesteps reversed a torch tensor with [::-1], which raises (torch has no negative-step slicing). Every scaffold inherited that from the template.

Replaced the reversal with torch.flip(..., dims=[0]) in:

• templates/scheduler/scheduling_TEMPLATE.py
• examples/scaffolded_scheduler/scheduling_ddpm_lite.py
• src/diffusers/schedulers/scheduling_euler_lite.py

step() stays TODO(engineer) — no sampler math. There is no scheduling_heun_lite.py in this kit.

Also added the missing dtype assertion on test_scheduling_euler_lite.py so TEST002 is 0 findings.

Convention gate

python3 tools/convention_check.py examples/scaffolded_scheduler src/diffusers/schedulers/scheduling_euler_lite.py
# 2 file(s) scanned, 0 findings

python3 tools/convention_check.py --all
# 18 file(s) scanned, 0 findings

Tests run

python3 -m unittest discover -s tests -t .
# Ran 52 tests in ~17s
# OK
# skipped_count 0

With torch 2.14.0+cpu and diffusers 0.40.0. Behavioral tests (test_timesteps_count, test_output_type, test_same_seed_same_output, test_config_roundtrip) passed on both ddpm_lite and euler_lite — they did not skip.

unittest log

_{To show artifacts inline, enable in settings.}