feat: real-time denoising step progress during video generation

[Sergio Gil Jiménez (jimeneztion)]

Summary

Previously the progress bar used time-based interpolation — it estimated how
long inference would take and animated accordingly. Progress was a guess and
could be wildly off.

This PR wires actual denoising step counts from the model to the frontend so
the bar reflects real work done.

Backend: tqdm interceptor (ltx_fast_video_pipeline.py)

The denoising loop in ltx_pipelines uses tqdm internally with no callback
hook. We patch tqdm in ltx_pipelines.utils.samplers at call time using a
context manager (_tqdm_progress_interceptor) that wraps each iteration and
fires a StepCallback(current_step, total_steps). The patch is scoped to the
generate call and restores the original tqdm in a finally block.

The distilled fast pipeline runs 11 total denoising steps (8 stage 1 + 3 stage 2),
exposed as constants.

Backend: progress reporting (video_generation_handler.py)

A _on_denoising_step callback maps each step to a progress percentage:
15 + 75 * (step / total), covering inference from 15% to 90%. It calls
update_progress("inference", pct, current_step, total_steps) so the frontend
receives both the percentage and the raw step counts.

Frontend (frontend/hooks/use-generation.ts)

• Replaces time-based inference interpolation with real currentStep /
  totalSteps from the backend SSE stream
• Shows "Generating... (5/11)" during inference
• Narrows loading-model interpolation from 15–95% to 5–12%
• Adds idle response filtering: ignores SSE events with empty phase or
  status === 'idle'
• Generalizes initial status message to "Loading model..."

How to test

1. pnpm dev
2. Start any video generation
3. Progress bar should show "Generating... (1/11)" → "Generating... (11/11)"
   as denoising steps complete — not a smooth timer animation
4. pnpm typecheck && pnpm backend:test