fix(deepinfra): reconcile reasoning controls

[rekram1-node]

Summary

• audit all 25 configured DeepInfra models using first-party model pages and the public catalog
• encode only positively established provider-wire reasoning controls
• represent DeepSeek V4 Non-think with reasoning.enabled toggle and Think High with reasoning_effort=high; omit Think Max because it is a special-system-prompt behavioral mode, not a supported reasoning-effort enum value
• map DeepSeek R1 0528 none to toggle and preserve exact low|medium|high effort through DeepInfra's documented R1 replacement chain
• omit reasoning options where exact provider-wire semantics remain unproven
• keep non-reasoning models free of reasoning options
• add a complete 25-model reasoning matrix regression test

Verified control matrix

• 25 models total: 21 reasoning, 4 non-reasoning
• toggle only: 10
• DeepSeek R1 0528: toggle plus effort low|medium|high
• DeepSeek V4: 2, toggle plus effort high
• GPT OSS: 2, effort low|medium|high
• fixed reasoning: 0
• token budget: 0
• non-reasoning/options omitted: 4

Unresolved omissions

• MiniMaxAI/MiniMax-M2.5
• Qwen/Qwen3.5-35B-A3B
• Qwen/Qwen3.5-397B-A17B
• Qwen/Qwen3.6-35B-A3B
• moonshotai/Kimi-K2.5
• zai-org/GLM-4.7-Flash

These remain reasoning=true with reasoning_options omitted because model-specific provider-wire configurability or fixed behavior is not positively established.

Validation

• bun validate
• bun test packages/core/test/deepinfra.test.ts (1 pass, 53 assertions)
• git diff --check

Audit notes

• no DeepInfra credential was available for authenticated behavior tests
• generic accepted API values and behavioral labels were not treated as effective model-wire controls
• PR fix(deepinfra): reconcile reasoning controls #2134 remains the only active DeepInfra provider PR