feat: add neuralwatt provider with 14 models

[monotykamary]

Summary

Adds Neuralwatt as an OpenAI-compatible inference provider with energy-aware GPU optimization.

Update: The Neuralwatt models API was updated and now returns accurate pricing, capabilities, and limits. Model IDs, pricing, and capabilities in this PR follow the upstream API faithfully — no manual patches needed (patch.json is now empty).

Provider Details

• API: https://api.neuralwatt.com/v1 (OpenAI-compatible)
• SDK: @ai-sdk/openai-compatible
• Auth: NEURALWATT_API_KEY
• Docs: https://portal.neuralwatt.com/docs

Models (14 total)

Reasoning models (with [interleaved] field = "reasoning_content")

Model ID	Name	Vision	Input $/M	Output $/M
moonshotai/Kimi-K2.5	Kimi K2.5	✅	$0.52	$2.59
moonshotai/Kimi-K2.6	Kimi K2.6	✅	$0.69	$3.22
kimi-k2.6-fast	Kimi K2.6 Fast	✅	$0.69	$3.22
zai-org/GLM-5.1-FP8	GLM 5.1 FP8	❌	$1.10	$3.60
MiniMaxAI/MiniMax-M2.5	MiniMax M2.5	❌	$0.35	$1.38
Qwen/Qwen3.5-397B-A17B-FP8	Qwen3.5 397B A17B FP8	❌	$0.69	$4.14
Qwen/Qwen3.6-35B-A3B	Qwen3.6 35B A3B	❌	$0.05	$0.10
openai/gpt-oss-20b	GPT OSS 20B	❌	$0.03	$0.16

Fast variants (optimized for speed, non-reasoning unless noted)

Model ID	Name	Vision	Input $/M	Output $/M
kimi-k2.5-fast	Kimi K2.5 Fast	✅	$0.52	$2.59
glm-5-fast	GLM 5 Fast	❌	$1.10	$3.60
glm-5.1-fast	GLM 5.1 Fast	❌	$1.10	$3.60
qwen3.5-397b-fast	Qwen3.5 397B Fast	❌	$0.69	$4.14
qwen3.6-35b-fast	Qwen3.6 35B Fast	❌	$0.05	$0.10

Other

Model ID	Name	Vision	Input $/M	Output $/M
mistralai/Devstral-Small-2-24B-Instruct-2512	Devstral Small 2 24B Instruct 2512	✅	$0.12	$0.35

Files Added

• providers/neuralwatt/provider.toml — Provider config
• providers/neuralwatt/logo.svg — Logo from official Neuralwatt favicon (adapted to currentColor)
• providers/neuralwatt/README.md — Provider documentation
• providers/neuralwatt/models/**/*.toml — 14 model definition files

Notes

• All model IDs, pricing, capabilities, and limits are sourced directly from the Neuralwatt API (/v1/models) — now fully accurate, no manual patches required
• Model IDs match the Neuralwatt API identifier format (some fast variants use bare names without org prefix, e.g. glm-5-fast)
• "Fast" variants are optimized for lower latency; kimi-k2.6-fast also supports reasoning
• Vision models support image input via OpenAI-compatible API
• Reasoning models use [interleaved] field = "reasoning_content" per the existing pattern
• Logo SVG adapted from the official favicon with fill="currentColor" for theme compatibility

Validation

• ✅ bun validate passes
• ✅ bun run build (web) succeeds