Relax triton pin from ==3.5.0 to >=3.4.0

[cennn]

Summary

• Change triton==3.5.0 to triton>=3.4.0 in requirements.txt
• The hard pin shadows the NGC-bundled pytorch-triton (3.4.0) when installing in PyTorch NGC containers, causing different Triton backend_hash values and numerically different compiled kernels — breaking compile-mode bit-exact reproducibility across images
• Since pyproject.toml declares no triton runtime dependency, this file is for development use only; relaxing the constraint lets the environment's existing triton be reused

Context

When MAGI-2's Dockerfile runs pip install -r requirements.txt before installing MagiCompiler, triton==3.5.0 overwrites the NGC base image's pytorch-triton 3.4.0. This produces a different backend_hash in Inductor-generated Triton kernels, leading to max ~5e-4 numerical differences in VAE encode that cascade to max ~1.78 divergence after DiT processing (chaotic amplification). The issue is absent in eager mode and only manifests in torch.compile mode.

Test plan

• Verify MagiCompiler imports and functions correctly with triton 3.4.0
• Rebuild MAGI-2 image without triton override, confirm compile-mode parity with Athena

[cennn]

Closing: the fix belongs in downstream Dockerfiles (Athena & MAGI-2) that strip the triton pin before installing, rather than relaxing it in MagiCompiler itself.