feat: add tensor parallelism awareness for multi-GPU model fitting

[michaeljabbour]

Summary

• supports_tp(tp_size) and valid_tp_sizes() methods on LlmModel for checking TP compatibility
• Heuristic head-count inference for Qwen, Llama, DeepSeek, Mistral, Gemma, Phi, MiniMax model families
• RunMode::TensorParallel with 0.9× speed penalty in fit analysis
• TUI: TP compatibility filter (Shift-T) cycles All → TP=2 → TP=3 → TP=4

Files changed (9)

File	Change
llmfit-core/src/models.rs	TP fields, methods, head-count inference, tests (+269)
llmfit-core/src/fit.rs	TensorParallel run mode, cluster-aware fit path (+54)
llmfit-core/src/hardware.rs	cluster_mode, cluster_node_count fields (+10)
llmfit-core/src/plan.rs	Test constructor updates (+4)
llmfit-tui/src/tui_app.rs	TpFilter enum, cycling, filter integration (+47)
llmfit-tui/src/tui_events.rs	Shift-T keybinding (+3)
llmfit-tui/src/tui_ui.rs	TP filter widget in search bar (+20)
llmfit-tui/src/serve_api.rs	TensorParallel in run_mode_code (+1)
llmfit-tui/src/main.rs	Test constructor updates (+6)

Test plan

• All existing tests pass (245 passed, 1 pre-existing failure unrelated to this PR)
• New TP unit tests: supports_tp, valid_tp_sizes, infer_heads_from_name for multiple model families
• Manual: launch TUI, press Shift-T to cycle TP filter, verify model list updates

Split from #181 — see closing comment for context on the split.

🤖 Generated with Claude Code

[AlexsJones]

what an awesome PR @michaeljabbour