fix(fast-path): use manifest max_field_id for new fields

[FANNG1]

Fixes #41

Problem

FastPathFragmentWriter computed the field ID for new columns from top-level schema fields:

next_fid = max(f.id() for f in self.lance_ds.lance_schema.fields()) + 1

lance_schema.fields() does not provide a complete high-water mark for field IDs. Nested fields, such as struct children, have their own Lance field IDs. Schema evolution can also leave previously used field IDs behind. If next_fid is based only on visible top-level fields, the new data file can be committed with a field ID that collides with an existing nested or historical field ID.

Example for a table with id (fid=0) and a struct meta (fid=1) with children meta.a (fid=2) and meta.b (fid=3):

Source	max fid	next_fid	Effect
Top-level fields only	1	2	Can collide with meta.a
Dataset max_field_id	3	4	Uses the manifest high-water mark

With the wrong next_fid, fragment metadata can associate the new file with the wrong schema field. Reads of the intended new column can then return null or wrong values.

Fix

Use Lance's manifest-level field ID high-water mark:

next_fid = self.lance_ds.max_field_id + 1

max_field_id reflects nested child fields and field IDs reserved by prior schema evolution, so it is the correct source for assigning new field IDs.

Test

Adds TestRegressions::test_next_fid_uses_manifest_max_field_id, which creates a dataset with a struct column, merges a new scalar column through the fast path, and verifies that the new column reads back with the expected values.