Set begin_index for flow-matching pipelines to avoid slow index_for_timestep

[kirillsst]

Summary

• Add scheduler.set_begin_index(0) before the denoising loop in all flow-matching pipelines that were missing it
• For img2img pipelines with strength-based schedule truncation (wan_video2video, ltx_condition), set the begin index to the appropriate start offset inside get_timesteps()
• Without begin_index, _init_step_index() falls back to index_for_timestep() which uses .nonzero() causing an expensive DtoH sync on every first scheduler.step() call
• This was already done in Flux and a few other pipelines but was missing across most flow-matching pipelines (56 files)

Details

The FlowMatchEulerDiscreteScheduler._init_step_index method has two paths:

if self.begin_index is None:
    self._step_index = self.index_for_timestep(timestep)  # slow: .nonzero()
else:
    self._step_index = self._begin_index  # fast: direct assignment

This PR ensures all flow-matching pipelines take the fast path, matching the pattern already established in pipeline_flux.py.

Fixes #9417

Test plan

• ruff check passes on all 56 modified files
• ruff format --check passes on all 56 modified files
• Existing pipeline tests should pass (no behavioral change, only performance)