Description:
This PR introduces better timing and profiling capabilities to the LTX2 generation pipeline to help identify performance bottlenecks.

Key Changes:

Detailed Timing: Added time.perf_counter() blocks and jax.block_until_ready() calls across the pipeline to accurately measure text encoding, connector passes, denoising steps, VAE decoding, and post-processing.

Multi-Pass Execution: Updated generate_ltx2.py to support a three-stage execution flow:

Warmup Pass: For JIT compilation.

Generation Pass: For actual output and standard timing.

Profiling Pass: (Optional) Captured via max_utils.Profiler for a subset of steps.

Enhanced Logging: Added a summary table for Load, Compile, and Inference times.

e.g.

==================================================
  TIMING SUMMARY
==================================================
  Load (checkpoint):     106.7s
  Compile:                81.5s
  ────────────────────────────────────────
  Inference:              16.4s
    Text Encoding:         3.6s
    Preparation:           0.0s
    Connectors:            0.0s
    Denoising:             7.2s
    Latent Upsampler:      0.0s
    Latent Processing:     0.6s
    Video VAE:             1.5s
    Video Post:            3.2s
    Audio VAE:             0.1s
    Vocoder:               0.2s
==================================================

Config Updates: Added skip_first_n_steps_for_profiler and profiler_steps to the LTX2 configuration.

Memory Management: Explicitly deletes large tensors (out, videos, audios) before the profiling run to prevent OOM.