feat(pkg-r): add ggsql visualization tool

[cpsievert]

Summary

R port of #201. Adds an opt-in visualize tool to querychat's R package, enabling LLM-generated data visualizations via ggsql. When enabled, the LLM can write ggsql queries (SQL + VISUALISE clause) that are executed and rendered as interactive charts inline in the chat.

Usage

library(querychat)

qc <- QueryChat$new(
  palmerpenguins::penguins,
  tools = c("filter", "query", "visualize")
)

qc$app()

The visualize tool is opt-in — include "visualize" in the tools vector alongside "filter" and/or "update" to enable it.

What it looks like

Charts render inline in the chat with a collapsible footer showing the ggsql query (with syntax highlighting), a copy button, and save options:

---

Key changes

• visualize tool (opt-in via tools = c("query", "visualize")): The LLM writes a full ggsql query, which is executed against the data source and rendered using ggsql's native Shiny bindings (ggsql::renderGgsql / ggsql::ggsqlOutput).
• Two-stage ggsql pipeline (querychat_viz.R): Uses DataSource for SQL execution (preserving database pushdown), then feeds results into ggsql's DuckDB reader for VISUALISE processing.
• Inline chart rendering: Charts render directly in the chat stream via Shiny's dynamic output system, with JS/CSS assets for the footer (show query toggle, save as PNG/SVG).
• Generated footer assets: pkg-r/inst/htmldep/viz.{js,css} are checked-in package assets, but the shared source/build logic for them already lives on main under js/ (see js/build.mjs). This PR is adding the R package's generated copies, not introducing a second implementation of that browser logic.
• PNG feedback to LLM: A static PNG is generated (best-effort, requires V8 + rsvg) and sent back to the LLM so it can see what it produced.
• collapsed parameter on querychat_query: Preparatory queries (e.g., inspecting data before visualizing) can start collapsed so the chart remains the focal point.
• Prompt restructuring: System and tool prompts were updated to match the Python package — visualization best practices, ggsql syntax reference, and conditional sections based on which tools are enabled.
• Bookmark support: Viz widget state is saved/restored during Shiny bookmarking.
• Viz state is internal only: No public accessor methods or reactive values for viz state — consistent with the Python implementation.

Test plan

• Unit tests for execute_ggsql, extract_visualise_table, has_layer_level_source
• Unit tests for tool_visualize_dashboard (creation, callback, error handling)
• Unit tests for truncate_error
• Unit tests for collapsed parameter on query tool
• Unit tests for viz prompt conditionals in system prompt
• Full test suite passes (622 tests, 0 failures)