Fix 15 correctness bugs from the full-codebase review

[vsdudakov]

A full review of the main branch surfaced 15 verified correctness bugs; this PR fixes all of them, each with regression tests placed in the topical test files.

Engine (Rust)

• ZoneInfo-aware datetime binding — with use_tz=True and any variable-offset IANA timezone (e.g. Europe/Berlin), every save/filter raised TypeError: expected a datetime without tzinfo. Aware datetimes whose fixed-offset extraction fails now normalize via astimezone(UTC).
• execute_many type corruption (Postgres) — parameter types were declared from the first row only; a later row with a different type (e.g. [[1], [2.5]] into DOUBLE PRECISION) wrote raw f64 bits into an INT8 parameter and silently stored a garbage value. Types are now unified across all rows (numeric widening INT8→FLOAT8→NUMERIC, clear error for incompatible mixes), with to_sql_checked hardening as a backstop.
• sync_fast_path event-loop deadlock (SQLite) — a sync-path query called block_on(pool.get()) on the loop thread; with a transaction pinning the pool's only connection the process hung forever. The sync path now uses a non-blocking checkout and falls back to the async path when the pool is busy.
• MSSQL bulk insert pk backfill — SCOPE_IDENTITY() - @@ROWCOUNT + 1 arithmetic assumed contiguous identity allocation, which SQL Server does not guarantee under concurrency, so objects could silently receive other rows' pks. The multi-row INSERT now returns real ids via OUTPUT INSERTED.<pk>.

ORM (Python)

• bulk_create discarded explicit pks on auto-increment models and overwrote obj.pk with generated ids; explicit pks are now honoured (wrapped in IDENTITY_INSERT where the dialect needs it) and mixed explicit/auto batches raise ValueError.
• get_or_create/update_or_create dropped the queryset's connection binding and filters: author.books.get_or_create(...) created orphan rows without the FK, and using_db(...) querysets selected on one database but inserted into the default. The create half now runs on the bound connection and inherits root-level equality filters (explicit kwargs/defaults win).
• Decode-plan race on reads — the snapshot-before-await invariant from 410ef33 (writes) now also covers _fetch/_fetch_annotated/_fetch_select_related, so a concurrent other-dialect compile can no longer swap decoders mid-hydration.
• annotate(n=...).order_by("-n").values_list(...) rendered ORDER BY "n" with no "n" in the SELECT — an error on PG/MySQL/MSSQL and silently wrong ordering on SQLite. Ordering annotations are kept in the SELECT (never leaked into results).
• MSSQL m2m deletes always failed (join-table FKs are NO ACTION there): Model.delete and QuerySet.delete now remove join rows first on non-cascading dialects; the bulk path materializes affected pks up front so join-dependent filters (e.g. annotate(n=Count(...)).filter(n__gte=2)) stay correct.
• Bare date in a DatetimeField was stored as date-only text on SQLite (refetched as str, breaking comparisons); it now widens to a tz-aware midnight datetime at assignment while __date lookups keep binding date-only values.
• through_fields orientation now follows Django's (owner_column, target_column) order — previously the columns were crossed, causing IntegrityErrors or silently swapped m2m rows. The forward_key/backward_key kwargs keep their documented (Tortoise-compatible) meaning; only the docstrings were wrong and are fixed.
• Lazy relation reads bypassed custom Meta.manager while prefetch respected it, leaking soft-deleted rows through author.books; both paths are now manager-scoped (add/remove/clear stay raw by design).
• Migrations: _field_source now emits pk=True for FK/O2O primary keys (previously a fresh DB got a table with no primary key plus a never-converging phantom AlterField); MySQL MODIFY COLUMN restates the column DEFAULT instead of silently dropping it; CreateModelIfNotExists is genuinely re-run safe on MySQL (indexes folded into the guarded CREATE TABLE) and MSSQL (sys.indexes existence guards).

Tests

Every fix has regression tests, placed in the topical test files; the old test_review_fixes.py grab-bag is redistributed the same way. Where feasible, tests were verified to fail against the pre-fix code.

• sqlite suite: 1250 passed
• postgres suite (live server): 1251 passed
• cargo test --lib: 53 passed; ruff + ty clean

Notes for reviewers

• MSSQL/MySQL-specific behavior is covered by rendered-SQL assertions and simulated capability flags (no live servers in this environment) — worth a pass against real servers before release.
• OUTPUT INSERTED fails on tables with triggers (documented in a comment; matches common ORM behavior).
• Known pre-existing gaps left out of scope, flagged in comments where relevant: MySQL standalone AddIndexIfNotExists remains unguarded (MySQL has no SQL-level guard), m2m prefetch still bypasses a custom target manager, and QuerySet.update(field=date(...)) still binds a raw date.