feat: provider-executed tools

[lgrammel]

Background

Many providers offer tools that are executed on their server and provide tool results, e.g. search and code execution tools. Omitting those outputs from the LLM history affects future calls (e.g. the LLM needs to know that it used a web search tool).

Summary

Introduces provider-executed tool calls, results and errors at all levels (language model, ai core, ui). Adds anthropic websearch support for full feature set.

Verification

End-to-end UI test: sources with anthropic websearch tool. render tool status and input.

Tasks

• spec updates
• send tool result from anthropic web search tool (stream)
• send tool error from anthropic web search tool (stream)
• send tool result from anthropic web search tool (generate)
• send tool error from anthropic web search tool (generate)
• generateText support for result with test
• generateText exclude provider side tool calls and results from roundtrip mgmt
• generateText support for error with test
• streamText support for result with test
• streamText support for error with test
• streamText exclude provider side tool calls and results from roundtrip mgmt
• add providerExecuted flag to tool calls, results, and errors
• UI extensions
• UI message mapping
• response message mapping
• simulate streaming middleware
• server side tool calls and results for openai file search
• server side tool calls and results for openai web search
• rm encrypted content from provider metadata
• add roundtrip generateText with server side tool example
• provider metadata on tool calls and results
• tool result error should have json value
• ui input streaming provider executed

Future Work

• implement google code execution tool
• implement full support for openai file and web search

Related Issues

Builds on #6807