[FIX] Make BIDSLayout.save idempotent when re-saving to the same path

[Leonard013]

Closes #1079.

Why

Calling layout.save(path) a second time against the same path raises:

sqlite3.OperationalError: table associations already exists

This bites users who re-run a script: the first run creates
<path>/layout_index.sqlite, and a second save() to the same path fails.
@effigies reproduced it in the issue thread and noted it wasn't obvious to fix.

Root cause

ConnectionManager.save_database (src/bids/layout/db.py) snapshots the index by
replaying sqlite3.Connection.iterdump() into sqlite3.connect(database_file).
iterdump() emits plain CREATE TABLE ... (no IF NOT EXISTS, no DROP), so
replaying into a database_file that already holds the schema (from a prior
save) collides on the first table — associations.

The reason a naive fix (deleting/truncating the destination first) does not
work is subtle, and is likely why this looked unfixable: with the default
replace_connection=True, the first save() re-points the layout's connection at
the saved file. A subsequent save() to the same path therefore has old_db
(source) and database_file (destination) aliasing the same file — clearing the
destination in place would destroy the source before iterdump() reads it
(confirmed: it produces an empty / no such table: layout_info database).

What

Dump into a fresh temporary file in the same directory, then atomically move it
into place:

temp_file = database_file.with_name(f'.{database_file.name}.tmp')
temp_file.unlink(missing_ok=True)
new_db = sqlite3.connect(str(temp_file))
with new_db:
    for line in old_db.iterdump():
        if line not in ('BEGIN;', 'COMMIT;'):
            new_db.execute(line)
    new_db.commit()
new_db.close()
temp_file.replace(database_file)

• The destination always starts empty, so CREATE TABLE never collides.
• The destination is only touched at the final replace(), so re-saving stays
  correct even when source and destination are the same file.
• Streaming via iterdump() is preserved; replace() is a same-directory atomic
  rename.

Tests

Added test_save_database_twice_same_path to src/bids/layout/tests/test_db.py:
saves a layout to the same path twice (default replace_connection=True,
exercising the same-file path) and asserts the re-saved database is complete and
reloadable.

• Before: fails with OperationalError: table associations already exists.
• After: passes.
• pytest src/bids/layout/tests/ → 193 passed, 71 skipped (network/submodule
  gated), 1 xfailed, 0 failed.

Caveats (flagging honestly)

• Windows, same-process same-file re-save: on the second same-file save,
  old_db still holds the destination open when replace() runs, which can raise
  PermissionError on Windows. This is not a regression (the current code
  already raises on the second save everywhere), and pybids CI is Linux/macOS only;
  first saves and cross-path saves are unaffected. Happy to guard it (close
  old_db before replace()) if you'd like it fully Windows-safe.
• Concurrent saves to the same path remain unsupported (unchanged from before).
• Happy to fold in polish if preferred: a process-unique temp name
  (tempfile.mkstemp(dir=...)), a try/finally to clean the temp on error, and/or
  a stronger completeness assertion in the test.

No CHANGELOG.rst edit — this project compiles the changelog from PR titles via
release-drafter.

---

Prepared with AI assistance (Claude Code): the reproduction, root-cause analysis,
fix, and test were AI-drafted; the reasoning and verification are documented above.