…1736)

The uncompressed branch of _compress_tiles allocated

    total_buf = bytearray(n_tiles * tile_bytes)
    mv = memoryview(total_buf)

at the top of the loop and never read either back. Tiles were still
built via tile_arr.tobytes() and appended to a list. The buffer was
roughly the size of the full uncompressed raster, so a marginal write
would OOM while holding both the dead buffer and the tile list.

Remove the dead allocation. Existing tile-build logic stays the same;
the round-trip is byte-for-byte unchanged.

Adds regression coverage in
test_writer_uncompressed_tiled_no_dead_alloc_1736.py for the exact
divisor, edge-tile padding, and multi-band paths.

Addresses two Copilot review comments on PR #1746:

* Update the module docstring's stale reference to ``_compress_tiles``
  (which doesn't exist) to point at the actual function name,
  ``xrspatial.geotiff._writer._write_tiled``.

* Add real teeth behind the "no dead alloc" test-module name. The
  three round-trip tests only covered correctness; nothing in the
  file would have failed if the ``bytearray(n_tiles * tile_bytes)``
  allocation were quietly restored.

Two new tests snapshot ``tracemalloc`` peak across a direct
``_write_tiled`` call (uncompressed branch, single- and 3-band 1024x1024
uint8 inputs) and assert peak stays under 1.5x raster bytes. The
current implementation lands at ~1.06-1.12x; the deleted bytearray
pushed peak to ~2.07x. The 1.5x cap reliably catches reintroduction
without flagging unrelated implementation drift.

The tracemalloc approach was preferred over renaming the tests/file
to "round-trip-only" because it preserves both the original intent
of the test name and gives future maintainers an automatic alarm if
the regression is reintroduced.