Support Lance Namespace in write_lance merge mode

[FANNG1]

Goal

Add well-defined namespace support for write_lance(mode="merge").

Scope

• Define behavior for a missing table, no new columns, join-key validation, and returned statistics.
• Reuse the namespace-aware merge implementation without duplicating fragile logic from Daft.
• Ensure merge commits honor namespace managed versioning.
• Add URI, directory-namespace, and Gravitino REST regression coverage.

Tracked by #51.

[FANNG1]

cc @universalmind303 @rchowell — want to lock down the API surface for merge before wiring up the Daft#7282 delegation.

Current state

• LanceDataSink explicitly rejects mode="merge" (the old sink path was never functional).
• Row-level merge already works via daft_lance.merge_columns_df(df, ...) / merge_columns(...), and both already support namespace addressing (table_id / namespace_impl / namespace_properties).
• Upstream DataFrame.write_lance still exposes mode="merge" + left_on/right_on and carries its own copy of the merge routing logic.

Question: for the Daft entry point, should write_lance(mode="merge") in daft-lance:

1. Support it — route mode="merge" to merge_columns_from_df and return a stats DataFrame like the other modes; or
2. Reject it — keep raising and point users to merge_columns_df, treating merge as a first-class daft-lance function rather than a write_lance mode.

I lean toward (2). write_lance and column-merge have genuinely different contracts — merge requires a fragment_id column and a join key, only ever adds columns (never rows), and requires the table to already exist. Folding that into a mode of an otherwise create/append/overwrite writer blurs the API without adding capability, since merge_columns_df already covers it.

So my plan for Daft#7282 would be: delegate create/append/overwrite to daft-lance's sink, and for mode="merge" raise a clear error pointing at merge_columns_df, deleting the duplicated merge logic from Daft. Objections before I go this route?