Set PYTORCH_CUDA_ALLOC_CONF before platform init

[Shan2L]

Fixes #1593

lightx2v/pipeline.py sets PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True via setdefault, but
lightx2v/__init__.py imports lightx2v_platform.set_ai_device before pipeline.py. Some platform
backends initialize the CUDA/HIP caching allocator during that import (on amd_rocm, importing aiter
does), and after that the allocator no longer reads the variable. As a result amd_rocm runs without
expandable segments, and MiniMax-H3 OOMs on 32 GB GPUs.

This PR applies the same setdefault at the top of lightx2v/__init__.py, before the platform import.
The existing line in pipeline.py is unchanged. An explicitly exported PYTORCH_CUDA_ALLOC_CONF still
wins. Platforms whose allocator does not read this variable are unaffected.

Test

• is_expandable of the first segment after import lightx2v: amd_rocm False -> True; cuda True (unchanged).
• MiniMax-H3 ref2av TP2 (768x1344, 362 frames, 3 steps, block offload) on 2x R9700, PLATFORM=amd_rocm, main:
  OOM at step 1 -> completes; 4 runs on 4 GPU pairs, 100.0-101.8 s/step (same as PLATFORM=cuda).
• ruff check / ruff format --check (v0.11.0) clean.

[Shan2L]

@helloyongyang @gushiqiao @llmc-reviewer could you take a look, or point me to the right reviewer? This is a 6-line fix for #1593: on amd_rocm the allocator is initialized before pipeline.py sets PYTORCH_CUDA_ALLOC_CONF, so MiniMax-H3 OOMs on 32 GB GPUs. CUDA behavior is unchanged and an exported value still wins.

It is one of two ROCm prerequisites (with #1596) for the gfx1201 attention backends in #1598 / #1599 (integration plan: ROCm/aiter#6292). Happy to adjust anything.