mcp: add data_url field to _table_catalog to enable mechanical table refresh

[StefanSteiner]

Summary

The _table_catalog source_url field currently stores a human-readable
page URL (e.g. https://ourworldindata.org/grapher/population). There is
no structured field for the actual data file URL, which means an automated
"refresh this table from its source" operation cannot be driven purely from
catalog metadata — it requires prose parsing from notes or manual
re-entry of the download URL.

Proposed change

Add a new data_url field to _table_catalog and expose it as a
parameter in set_table_metadata. This is additive and non-breaking:
source_url continues to hold the human-readable page/reference URL,
source_description is unchanged, and data_url carries the machine-
actionable download URL.

Example for an OWID table:

Field	Value
source_url	https://ourworldindata.org/grapher/population (chart page — unchanged)
source_description	"OWID historical population, 10 000 BCE – 2023…" (unchanged)
data_url	https://ourworldindata.org/grapher/population.csv?v=1&csvType=full&useColumnShortNames=true ← new

Why this matters — mechanical refresh

With data_url and the existing load_params (which already records
mode, schema, merge_key, database, etc.), a refresh becomes fully
mechanical with no prose parsing:

1. SELECT data_url, load_params
   FROM _table_catalog
   WHERE table_name = 'owid_population'

2. Infer format from data_url extension (.csv / .parquet / .json / etc.)

3. Download data_url → temp file

4. load_file(path=<temp_file>, **load_params)

This would enable:

• A refresh_table tool that re-ingests any table from its source in one
  call.
• Bulk refresh of all tables that have a data_url set:
  SELECT table_name FROM _table_catalog WHERE data_url IS NOT NULL.
• Scheduled refresh without any human-maintained refresh scripts.

Format inference rule

Infer from the last path segment's extension before any query string
— the same rule load_file already uses for local paths:

URL ending	Format
.csv	CSV
.parquet	Parquet
.json / .jsonl	JSON / NDJSON
.arrow / .ipc	Arrow IPC

For URLs where the extension is absent or ambiguous (e.g. a presigned S3
URL with a UUID key), a source_format hint in load_params or an
additional data_format field could carry the override.

Related issue

This pairs with the rename-drops-metadata issue: both are about making
_table_catalog fully machine-actionable rather than just human-readable.
A refresh_table tool (future) would depend on both fixes — metadata
surviving renames, and a structured data_url to fetch from.

Environment

• hyper-rust-api version: 0.2.3.re93d08d2
• Observed while loading 17 OWID datasets into the persistent database
  and noting that refresh requires prose parsing of notes.