feat(databricks): create tables with IF NOT EXISTS and reconcile columns for concurrent loads

[francescomucio]

Closes #4135

Problem

When several dlt runs load into the same Databricks table concurrently and the table doesn't exist yet, every run reads storage, sees the table missing (generate_alter = False), and emits a plain CREATE TABLE. The first commit wins; the rest fail with "table already exists". Hiding behind that crash is a second bug: a run carrying an extra column would see its create become a no-op and then fail when its load writes a column that was never added.

Change (Databricks only)

• CREATE TABLE IF NOT EXISTS for all tables. _make_create_table is overridden to drop the "dlt system tables only" restriction (still gated by the existing supports_create_table_if_not_exists capability). Covers both the base CREATE path and the custom cluster/partition/tblproperties path.
• Column reconciliation after the create. Databricks has no ADD COLUMN IF NOT EXISTS, so _execute_schema_update_sql re-reads the destination after the (possibly no-op) create and issues ALTER TABLE … ADD COLUMN for any columns still missing, tolerating the FIELDS_ALREADY_EXISTS error raised when a concurrent run adds the same column first.

Net effect: identical-schema parallel loads stop racing, and divergent-schema loads converge.

Notes

• The reconciliation read runs only inside update_stored_schema (i.e. when the schema hash actually changes), not on every load.
• Reconciled columns may not appear in the per-load-package applied_update diff; this is informational only — the data write maps columns by name against the full stored schema, so the load is unaffected.

Tests

Added to tests/load/databricks/test_databricks_table_builder.py (no live connection needed):

• CREATE TABLE IF NOT EXISTS emitted on both create paths.
• Reconciliation adds a column a concurrent create missed.
• FIELDS_ALREADY_EXISTS is tolerated; other errors re-raise.