Support alpha cumulative product using shifted sigmas for Flux

[rockerBOO]

In many diffusion models, xsigma² = (1-α)/α where α is the cumulative product of alphas. So we can derive alphas_cumprod from sigmas

The sigmas are shifted in initialization of the noise_scheduler so we can use the shifted sigmas, unless the scheduler is using a dynamic shifting (based on resolution). I am not detecting if the scheduler is using dynamic shifting but we can add this clarification.

Noise scheduler's from Diffusers have index_to_timestep which we can use to better align with future timestep alignment.

Supporting timesteps in ranges of 0 to 1 and 1 to 1000.

Support in Flux network training:

• min SNR gamma --min_snr_gamma 5.0
• debiased estimation --debiased_estimation

Flux huber loss using snr:

• huber c SNR --huber_schedule snr

v-pred in this PR currently but not sure if it is appropriate for flux.

---

Testing of the parameters and making sure regression for other models due to changing the timestep process to account for finding alpha_cumprod where needed. We might want to preload alpha_cumprod to reduce the complexity where appropriate (maybe not appropriate if using dynamic shifting).

Related #1980

[rockerBOO]

• Updated loss modification functions to take image_size for dynamically shifting timesteps (timestep_sampling == "flux_shift").
• Added tests for loss modification in custom train functions
• Updated FlowMatchEulerDiscreteScheduler
  • Added FlowMatchEulerDiscreteScheduler.get_snr_for_timestep to get snr of a timestep, accounting for dynamic timestep shifts.
• Updated flux noise scheduler setup to account for dynamic timestep shifts.

Need to do some tests and maybe cleanup prepare function (maybe not need to be duplicated).