Fix nan sigmas when num_inference_steps=1 with shift_terminal

[promptsmith1990]

What does this PR do?

Fixes #14411.

FlowMatchEulerDiscreteScheduler.stretch_shift_to_terminal computes one_minus_z[-1] / scale_factor, where scale_factor is itself derived from one_minus_z[-1]. With num_inference_steps=1 the single sigma in the schedule is both the first and last point, and before any terminal stretching it is exactly 1.0 (pure noise), so one_minus_z[-1] is 0 and the division degenerates to 0/0, producing nan sigmas and timesteps. Any pipeline configured with shift_terminal (LTX, LTX2, ...) run with a single denoising step hits this.

While investigating, I found the same unguarded if self.config.shift_terminal: call-site pattern copy-pasted into two other schedulers that share the identical stretch_shift_to_terminal implementation — UniPCMultistepScheduler's use_flow_sigmas branch, and FlowMatchLCMScheduler — and both reproduce the same nan crash under the same conditions, so this PR fixes all three call sites the same way: skip the stretch when there's only one point in the schedule to stretch. Multi-step schedules are unaffected and still terminate at shift_terminal as before (covered by a regression test).

Testing

pytest tests/schedulers/test_scheduler_flow_match_euler_discrete.py -v
pytest tests/schedulers/test_scheduler_unipc.py -k flow_sigmas_single_step -v

FlowMatchEulerDiscreteScheduler had no dedicated test file at all despite being used across ~20 pipeline test files, so I added tests/schedulers/test_scheduler_flow_match_euler_discrete.py covering: multi-step still reaches shift_terminal (regression guard against the fix disabling stretching generally), single-step no longer produces nan sigmas/timesteps, and a single scheduler.step() call runs to completion without nan output. A matching regression test was added to test_scheduler_unipc.py for the use_flow_sigmas branch.

Also ran utils/check_copies.py and ruff check/ruff format on the changed files (clean) to confirm the stretch_shift_to_terminal # Copied from block itself was left untouched — only the calling guard changed at each site.

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? If so:
  • Did you read the Coding with AI agents guide?
  • Did you run the self-review skill on the diff?
  • Did you share the final self-review notes in the PR description or a comment?
• Did you read the contributor guideline?
• Was this discussed/approved via a GitHub issue or the forum? Linked to Division by 0 bug when only one denoising step #14411, not yet acked by a maintainer.
• Did you write any new necessary tests?

AI disclosure

This PR was prepared with AI assistance (Claude Code). The assistant found the bug while investigating issue #14411, traced the same defect into the two other schedulers, wrote the fix and tests, and ran the verification shown above. I (the human submitter) reviewed the diff and the verification output before submitting, and take responsibility for this contribution.

Self-review notes (against .ai/references/review-rules.md)

Blocking issues: none found.

Non-blocking (left for actual review): the new test_scheduler_flow_match_euler_discrete.py uses a plain unittest.TestCase rather than SchedulerCommonTest, since this scheduler currently has no dedicated test file at all and adopting the full common-test harness (dummy sample generation, forward-pass parity checks, etc.) for the first time here is a separate, larger effort than this bug fix.

Verdict: READY.

Who can review?

@yiyixuxu @dg845 — schedulers