Add regression tests for FlowMatchEuler, DDIM guard, and LoRA state dict utils

[cursor]

Risky behavior now covered

FlowMatchEulerDiscreteScheduler

The default scheduler for SD3, Flux, Wan, Qwen-Image, and many other pipelines had no dedicated scheduler test file. Pipeline tests only smoke-test it as a black box. New contract tests lock down:

• Mutually exclusive sigma schedule config validation
• Dynamic shifting requiring mu
• Custom sigmas/timesteps length validation
• Integer timestep index rejection in step() (a common pipeline footgun)
• Stochastic sampling branch

DDIMScheduler set_timesteps guard

An eval branch demonstrated that inverting the num_inference_steps > num_train_timesteps comparison silently breaks all valid inference schedules. New tests assert the guard rejects over-limit steps and accepts the boundary value.

LoRA state dict conversion (state_dict_utils)

Every LoRA load path goes through convert_state_dict_to_peft / convert_state_dict_to_diffusers / convert_unet_state_dict_to_peft, but these had zero direct unit tests. New tests verify key remapping for diffusers-old, PEFT, and UNet formats, plus error paths for unrecognized state dicts.

Test files added/updated

• Added tests/schedulers/test_scheduler_flow_match_euler_discrete.py (11 tests)
• Updated tests/schedulers/test_scheduler_ddim.py (2 tests)
• Added tests/others/test_state_dict_utils.py (6 tests)

Why these tests materially reduce regression risk

1. High blast radius: FlowMatchEuler and LoRA conversion are shared infrastructure used by dozens of pipelines.
2. Silent failure modes: Inverted comparison guards and incorrect LoRA key remapping produce wrong results without raising errors.
3. Validation logic: The new tests target explicit ValueError paths that are easy to break during refactors but invisible to pipeline smoke tests.

Validation

All 19 new tests pass locally:

python3 -m pytest tests/schedulers/test_scheduler_flow_match_euler_discrete.py \
  tests/schedulers/test_scheduler_ddim.py::DDIMSchedulerTest::test_set_timesteps_num_inference_steps_exceeds_train_timesteps_raises \
  tests/schedulers/test_scheduler_ddim.py::DDIMSchedulerTest::test_set_timesteps_num_inference_steps_at_limit_succeeds \
  tests/others/test_state_dict_utils.py -v