#535 Postgres style indexes

[Benjamin-Knight]

Based off @axellpadilla original work, been using myself for a couple of weeks so appears stable, I was about to base this on the #710 work due to the transactions and online/resumable indices but instead have made it independent of that and just moved them out of the transaction completely.

Postgres-style indexes model config for SQL Server

Closes #535.

Adds a declarative indexes config (modelled on dbt-postgres) so indexes can
be defined in a model/seed/snapshot config block instead of hand-written
post-hooks. dbt owns the full lifecycle: it creates them on build,
reconciles them against the config on subsequent runs, and never touches
indexes it didn't create.

Usage

  {{ config(
      materialized="table",
      as_columnstore=False,
      indexes=[
        {"columns": ["id"], "type": "clustered", "unique": True},
        {"columns": ["created_at", {"column": "status", "desc": True}],
         "included_columns": ["amount"], "where": "status is not null"},
        {"columns": ["lookup_key"], "build_options": {"online": True}},
      ]
  ) }}

What's included

Config surface — most of CREATE INDEX is expressible per entry:

• columns (ascending, or {"column": name, "desc": true}), included_columns,
  unique
• type: clustered / nonclustered / columnstore
• data_compression (row/page, or columnstore/columnstore_archive), where
  (filtered indexes), fillfactor, pad_index, ignore_dup_key,
  optimize_for_sequential_key
• build_options for build-time-only settings (online, resumable, maxdop,
  sort_in_tempdb, lock options, …) — these affect how the index is built, not
  what it is, so they never trigger a rebuild
• Cross-config validation (e.g. at most one clustered index; clustered
  rowstore vs as_columnstore conflict), surfaced on the first build rather
  than mid-run

Deterministic naming & idempotency — managed index names are a hash of the
full definition (dbt_idx_ prefix). Name equality ⇔ definition equality, so
creation is idempotent and an unchanged definition is never rebuilt.

Reconciliation — on the paths where a relation persists across runs
(incremental non-full-refresh, DML table refresh, snapshot updates), the
index set is converged on the config in one atomic batch:
dropped-then-recreated definitions, with no window where a replacement index
is missing. Constraint-backing indexes, legacy post-hook
(clustered_/nonclustered_) indexes, and the as_columnstore CCI are never
dropped. drop_unmanaged_indexes (false (default) / warn / true) controls
treatment of indexes dbt didn't create.

Transaction-safe — the reconcile batch is @@TRANCOUNT-aware: it owns its own
transaction under autocommit (the default) and joins an existing one if dbt
is managing the model transaction, so it stays correct either way.
online/resumable indexes are built in the materialization's post-commit
phase, since SQL Server forbids those operations inside a transaction; these
builds are idempotent but not atomic with the model.

Lock-free introspection — index introspection reads the catalog with NOLOCK
on every sys view, so it doesn't queue behind concurrent index DDL.

Limitations

• Minimum supported SQL Server version is 2017 (STRING_AGG ... WITHIN
  GROUP).
• Partitioning, XML_COMPRESSION, and ordered columnstore are not yet
  expressible.
• online/resumable build options are not supported on seed models when dbt
  manages the transaction — dbt-core's seed materialization isn't
  adapter-owned and has no post-commit hook (documented in the changelog).

[LuuOW]

Technical audit: Implementation verified for system consistency and engineering integrity.