[Feature] Enable nd tiling workaround for dynamic shapes

[themistbeforedawn]

🗂️ PR Category

• ✨ New Feature
• 🚀 Optimization (performance, memory, etc.)
• 💥 Breaking Change
• 🐛 Bug Fix
• 🛠️ Development / Refactoring
• 📚 Documentation
• 🧹 Chore (Dependencies, CI/CD, Configuration, etc.)
• 🧪 Testing

📝 Description

Problem

On PyTorch < 2.11.0, Inductor's coalesce tiling analysis bails out on symbolic
numels under dynamic shapes, degrading transpose/permute/channels-last pointwise
kernels to untiled Grid1D — a big hit on memory-bound workloads like VAE decode.

Fix

Auto-apply a Triton ND-tiling workaround for dynamic-shape compilation:
prefer_nd_tiling=True, max_tiles=3, tile_reductions=True.

Performance (WAN 2.2 VAE decode, 540p, dynamic H/W)

OFF	ON	Speedup
770 ms	530 ms	~1.45x

Gating

• Dynamic shapes only — static shapes already tile well; forcing it on is a net loss.
• Version gated — auto-disabled on PyTorch >= 2.11.0 (fixed upstream).
• Override — compile_config.enable_dynamic_nd_tiling (settable via the MAGI_COMPILE_ENABLE_DYNAMIC_ND_TILING env var) → auto.

Tests

• Logic (28 cases): dynamic/version detection, explicit-config override precedence, env-var-drives-config, config injection.
• Perf: weight-free VAE-decode-like workload (3D conv + upsample), dynamic H/W; ~1.35x ON vs OFF on H100 (asserts >= 1.20x on calibrated GPUs).