docs(python/callbacks): refresh BigQuery callback page for shipped API (#3809)

Refreshes
[`/oss/python/integrations/callbacks/google_bigquery`](https://docs.langchain.com/oss/python/integrations/callbacks/google_bigquery)
so it matches the ``BigQueryCallbackHandler`` API as it actually shipped
in
[langchain-ai/langchain-google#1721](langchain-ai/langchain-google#1721)
and
[langchain-ai/langchain-google#1740](langchain-ai/langchain-google#1740).
Mirrors the README in the upstream package.

## Why

Following along the published Python BigQuery callback page today drops
users into stale APIs:

- Event names: `GRAPH_START`/`GRAPH_END`/`GRAPH_ERROR` and
`NODE_STARTING`/`NODE_COMPLETED`/`NODE_ERROR` (renamed to `INVOCATION_*`
and `AGENT_*` in #1721 to match Google ADK's
`BigQueryAgentAnalyticsPlugin`).
- Default table: `agent_events_v2` (renamed to `agent_events` in #1740).
- Six new `BigQueryLoggerConfig` options shipped in #1721
(`custom_tags`, `log_session_metadata`, `content_formatter`,
`auto_schema_upgrade`, `create_views`, `view_prefix`) plus
`skip_internal_chain_events` — none of which appeared on the page.
- Auto-created `v_*` analytics views, sub-agent attribution, `flush()`,
`model_version` / `usage_metadata` / `cache_metadata` / `llm_config` /
`tools` enrichment — all ADK-parity features missing from the page.
- One stale `USER_MESSAGE_RECEIVED` reference (an ADK-only event the
LangChain handler doesn't emit).

## What changed

| Section | Change |
|---|---|
| **Key features** | Rewritten to call out auto-views, auto schema
upgrade, sub-agent attribution, rich LLM telemetry,
`skip_internal_chain_events`, and `flush()`. |
| **Use with LangGraph agent** | Code sample updated: drops the explicit
`table_id="agent_events_v2"` (now the default), shows `custom_tags`,
demonstrates `flush(timeout=5.0)` between requests. |
| **Configuration options** | Extended with the 7 new options.
`table_id` default updated to `agent_events`. |
| **NEW: Auto-created analytics views** | Lists each `v_*` view and the
typed columns it exposes. |
| **NEW: Auto schema upgrade** | Documents the additive `ALTER TABLE ADD
COLUMN` behavior gated by the `langchain_bq_schema_version` table label.
|
| **NEW: Sub-agent attribution** | Documents the `agent → langgraph_node
→ checkpoint_ns → graph_name` fallback chain. |
| **Event types and payloads** | LLM_RESPONSE attributes now mention
`usage`, `model_version`, `usage_metadata`, `cache_metadata`.
LLM_REQUEST mentions `llm_config` + `tools`. New AGENT_* / INVOCATION_*
table. Tool usage table fixed to show the actual `{tool, input/result}`
JSON shape. |
| **Advanced SQL examples** | Sweep stale event names.
Sessionize-conversation view fixed (no more `USER_MESSAGE_RECEIVED`).
Tool-name path fixed (`$.tool` not `$.tool_name` in content). |
| **LangGraph integration** | Event-type table updated to `AGENT_*` /
`INVOCATION_*`. Code sample uses the renamed events. |

Total: **172 insertions, 53 deletions** in one file. Net 25 stale tokens
swept (`GRAPH_*`, `NODE_*`, `agent_events_v2`, `USER_MESSAGE_RECEIVED`).

## Test plan

- [x] Page renders as valid MDX (no broken frontmatter, code fences, or
tables).
- [x] All `JSON_VALUE` paths in the SQL examples match the actual JSON
shape the handler emits (verified against `_build_content` in
`bigquery_callback.py`).
- [x] All `event_type` literals in code/SQL examples are types the
handler actually emits.
- [x] Zero remaining stale references (`grep -cE
"GRAPH_(START|END|ERROR)|NODE_(STARTING|COMPLETED|ERROR)|agent_events_v2|USER_MESSAGE_RECEIVED"`
returns 0).

Author: caohy1988