Add CPU-first inference with automatic SDPA fallback

[Lakshya717]

Summary

Makes LingBot-Map run on CPU-first setups without FlashInfer, so inference works on machines with no capable GPU or missing CUDA wheels.

Why

The repo assumed CUDA + FlashInfer by default. On this dev machine (i5-13420H, 13 GB RAM, RTX 2050 4 GB), FlashInfer was unavailable and the ~4.4 GB checkpoint cannot fit in VRAM. The goal was reliability over speed — run anywhere, GPU optional.

What changed

• lingbot_map/utils/device.py — device/dtype resolution, autocast helper, FlashInfer probe, SDPA vs FlashInfer backend selection
• demo.py — --device (cpu default), auto SDPA when FlashInfer missing, device-aware autocast, images stay on CPU, --offload_to_cpu default on
• aggregator/stream.py — soft FlashInfer → SDPA fallback at build time (no crash)
• gct_base.py — device-aware autocast in prediction heads
• Benchmark — CPU/SDPA-safe defaults in method + YAML configs
• geometry.py — replace hard .cuda() with .to(device)
• README — appended CPU/low-spec install and usage notes (CUDA path unchanged)
• Profiling scripts — clearer CUDA-only errors

FlashInfer remains an optional CUDA fast path; SDPA is the portable default.

Status

Successful — CPU inference path works without FlashInfer. Verified locally: FlashInfer import fails → SDPA selected; --device auto correctly falls back to CPU on low VRAM (~3.2 GiB free).

Local results (this machine)

• Hardware: i5-13420H · 13 GB RAM · RTX 2050 4 GB (WSL2)
• Run: demo.py --device cpu (SDPA, no FlashInfer)
• Output: https://drive.google.com/drive/folders/1pOxZ3_mr8Lrjsck6V_GWY54kho3nSiSz?usp=sharing

Follow-up (not in this PR)

Further refinement is needed for hybrid / partial offload — e.g. weights on CPU with selective GPU compute when VRAM allows, or smarter auto-switching between CPU and CUDA within a single run. Current scope is CPU-only or explicit CUDA, not layered weight offload.