[codex] Validate model tool calls

[gold-silver-copper]

Summary

• Preserve the executable Rig tool names advertised for each completion turn.
• Reject unknown model-emitted tool calls with PromptError::UnknownToolCall before hooks, tool execution, or tool-result feedback in both prompt paths.
• Surface Gemini malformed/unexpected tool-call finish reasons as completion errors instead of empty successful streamed turns.
• Allow .tool(...).hook(...).build() and cover it with a compile regression test.

Root Cause

Rig trusted structured tool calls too late in the prompt loop. If a provider/model emitted a tool name Rig had not advertised for the current request, the call could proceed into normal hook/tool handling and corrupt the multi-turn state. For Gemini streaming this could show up later as an empty final response warning rather than the real protocol error.

Validation

• cargo fmt
• cargo check -p rig-core
• cargo test -p rig-core agent::prompt_request::
• cargo test -p rig-core providers::gemini::
• cargo test -p rig-core hook_can_be_set_after_tool_configuration