fix: split bulk inserts that exceed SQLite's bind parameter limit

[wtsnz]

The problem

SQLite takes at most 32,766 bind parameters per statement (SQLITE_MAX_VARIABLE_NUMBER). A bulk create inserts each batch in one statement, which needs rows × fields parameters. So a large batch, or a wide resource even at the default batch_size of 100, fails:

inputs = for i <- 1..15_000, do: %{title: "title#{i}"}
Ash.bulk_create(inputs, Post, :create, batch_size: 15_000, return_errors?: true)
# => error: "** (Exqlite.Error) variable number must be between ?1 and ?32766"

The fix

bulk_create/3 now inserts a batch in chunks that fit:

• the rows per chunk are the 32,766-parameter limit, less 1,000 of headroom for the ON CONFLICT clause's own parameters, divided by the widest row;
• the chunks' counts and returned rows are combined, in input order, since the returned rows are paired with the changesets by position.

A batch that fits is still a single statement. Ash runs each batch in a transaction, so a split batch stays atomic.

Tests

test/bulk_create_test.exs bulk creates 15,000 posts in one batch, checking that it succeeds and that every row is stored.

• On main, it fails as above.
• With the fix, it passes. Against AshSQL main (ASH_SQL_VERSION=local), the whole suite passes (279 tests).

Separately, AshSqlite main's own lockfile pins AshSQL 0.7.5. That predates the grouped aggregate strategy ash-project#232 relies on (ash-project/ash_sql#264). So with the locked dependencies, 89 aggregate tests fail on main itself, whatever this change does. They pass once AshSQL is released with #264 and the lock is bumped.

Found by

The data-layer conformance suite (wtsnz/ash#7), gap bind-parameter-limit, scenario large.bulk_create_parameters. With this branch, it passes. The same bug in AshPostgres, where it also drops the connection, is fixed in wtsnz/ash_postgres#8.