Slight performance improvement to Euler, EDMEuler, FlowMatchHeun, KDPM2Ancestral

[hlky]

What does this PR do?

This PR provides a slight performance improvement to Euler, EDMEuler, FlowMatchHeun by moving creation of randn tensor noise to inside the gamma > 0 condition.

gamma, calculated as min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0 is generally 0.0 as s_churn, s_tmin and s_tmax are typically not changed from their default values of 0.0, 0.0 and inf respectively.

Similarly in KDPM2Ancestral we move creation of noise to inside the 2nd order path.

Who can review?

@yiyixuxu

[vladmandic]

@hlky love the attention you're giving schedulers.

btw, a bit off-topic, but though of mentioning it here:
this is executed on each step - and its definitely causing torch sync:

sample = sample.to(torch.float32)

[hlky]

Thanks! Looks like that cast and the mentioned precision issues comes from rescale_betas_zero_snr so perhaps the cast could be moved under if self.config.rescale_betas_zero_snr, or a different value like in Comfy and Forge could be used and the cast removed.

[vladmandic]

rescale_betas_zero_snr

that options is so rarely used that moving cast only if its true should be more than ok.

[yiyixuxu]

thanks!

[yiyixuxu]

some of the tests failed as expected because the random seeds are used differently now
can we:

1. make a simple slow test that running a pipeline with each of these scheduler we changed and make sure the output is expected (even though it might be different from the main even with the same seed)
2. update the tests!

[hlky]

1. Do we want new slow tests like

   diffusers/tests/pipelines/stable_diffusion/test_stable_diffusion.py

   Lines 919 to 931 in e16fd93

   	def test_stable_diffusion_lms(self):
   	sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", safety_checker=None)
   	sd_pipe.scheduler = LMSDiscreteScheduler.from_config(sd_pipe.scheduler.config)
   	sd_pipe = sd_pipe.to(torch_device)
   	sd_pipe.set_progress_bar_config(disable=None)
   	inputs = self.get_inputs(torch_device)
   	image = sd_pipe(**inputs).images
   	image_slice = image[0, -3:, -3:, -1].flatten()
   	assert image.shape == (1, 512, 512, 3)
   	expected_slice = np.array([0.10542, 0.09620, 0.07332, 0.09015, 0.09382, 0.07597, 0.08496, 0.07806, 0.06455])
   	assert np.abs(image_slice - expected_slice).max() < 3e-3

   
   or just review the output to check it's still reasonable?
2. I've updated the values in KDPM2AncestralDiscreteSchedulerTest

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.