Implement maintainer-owned multi-GPU fit simulation

[Andyyyy64]

Multi-GPU support is now requested in #65 and #110, with deployment strategy details in #52. #84 also has a useful first pass, but current main has moved enough that I want to rework this in smaller maintainer-owned steps.

First-pass scope:

• parse repeated and comma-separated --gpu values, including forms like 2x RTX 4090
• represent heterogeneous GPUs without treating them as one perfect pooled device
• use conservative VRAM fit math for multi-GPU setups
• keep speed estimates low-confidence unless backend split assumptions are explicit
• show warnings that explain split and heterogeneous-GPU assumptions

Out of scope for the first pass:

• exact TP/DP throughput prediction
• PCIe lane and NVLink modeling
• backend-specific tensor-split tuning

References:

• Feature Request: Multi-GPU simulation via --gpu flag (heterogeneous setups) #65
• Multi-GPU Deployment Configuration and Optimal Parallel Strategy #52
• Feature/multi gpu #84
• support multiple GPUS Parallelism, e.g. TP or DP #110

The implementation should preserve existing single-GPU behavior and include regression tests for homogeneous GPUs, heterogeneous GPUs, and partial-offload cases.