Add TensorX provider

[jeremybower]

Summary

• Adds the tensorx provider (EU-sovereign inference, OpenAI-compatible API at https://api.tensorx.ai/v1).
• Adds 24 models with pricing (input/cache_read/output per 1M tokens) sourced from the TensorX pricing page.
• Reuses provider-agnostic model metadata via base_model for 18 models; full definitions for the 5 without existing metadata (deepseek-v3.2, deepseek-chat-v3.1, deepseek-r1-0528, qwen3-235b-a22b-2507, qwen3-vl-235b-a22b-instruct) plus the qwen3-embedding-8b embedding model.

Reasoning options

• deepseek-v4-pro / deepseek-v4-flash: toggle + effort ("max") + interleaved (reasoning_content), per the TensorX Chat Completions docs which document extra_body.chat_template_kwargs.thinking and extra_body.reasoning_effort.
• All other reasoning models: reasoning_options = [] — the docs note "Behaviour on other models may differ," so no control is claimed without verification.

Notes

• Context overrides applied where TensorX publishes a different window than upstream metadata (e.g. minimax-m3 → 1M, minimax-m2.5 → 192K, GLM-5 family → 198K).
• qwen3-vl-235b-a22b-instruct: attachment = true (vision), reasoning = false per repo convention for instruct variants.
• deepseek-chat-v3.1: reasoning = false (non-thinking chat variant).

Validation

• bun validate (exit 0)
• cd packages/web && bun run build (exit 0)
• Generated output confirms 24 models resolve with correct pricing, limits, and reasoning options.