feat: add conservative multi-GPU simulation

[Andyyyy64]

summary

• accepts repeated, comma-separated, and count shorthand --gpu specs, including 2x RTX 4090
• computes a conservative effective VRAM budget for multi-GPU fit checks instead of treating raw VRAM as one perfect pool
• keeps multi-GPU speed low confidence and applies a conservative speed factor
• exposes multi-GPU metadata in JSON output

scope

This is the fit simulation pass for #112. It does not try to model exact tensor parallel or data parallel throughput, PCIe lane layout, NVLink, NCCL/RCCL behavior, or backend-specific tensor splits. Those belong in #52.

Fixes #112.
Refs #65, #52, #84, #110.

hardware feedback

@cobra91 @theodufort @Honghe, if you still have access to the multi-GPU systems from #65, #84, or #110, could you try this branch and paste the hardware panel plus the top results? I mainly want to see whether the detected GPU list, effective VRAM warning, and top recommendations look sane on real hardware.

Suggested commands:

uvx --from "git+https://github.com/Andyyyy64/whichllm.git@feature/multi-gpu-fit-simulation" whichllm hardware
uvx --from "git+https://github.com/Andyyyy64/whichllm.git@feature/multi-gpu-fit-simulation" whichllm --status --top 5 --evidence any

@0xDE57, I kept PCIe lane and interconnect modeling out of this PR. The fit math is conservative, but topology-aware deployment strategy should stay in #52.

validation

• uvx ruff check .
• uvx ruff format --check .
• uv run python -m compileall -q src tests
• uv run pytest (350 passed)
• manual CLI checks for single GPU, repeated --gpu, comma-separated specs, count shorthand, invalid --vram, JSON output, 2x RTX 4090, mixed RTX 4090 + RTX 3060, and 4x RTX 4090