Add tests for the streaming reproject fallback path (#3101)

[brendancol]

Closes #3101.

• reproject()'s streaming fallback (_reproject_streaming, _process_tile_batch, _parse_max_memory) only runs when the source is over 512 MB and dask is not importable, so nothing in CI ever executed it. The new test file calls the helpers directly on small rasters and checks the output against the in-memory numpy path on the same grid: multi-tile threaded, serial (max_memory=1 byte), single tile, and nearest resampling with NaN cells. _parse_max_memory gets 10 unit cases (int, float, KB/MB/GB/TB suffixes, lowercase with whitespace, bare numeric string, None default).
• The 3-D source crash in the streaming assembly is pinned with a strict xfail referencing _reproject_streaming crashes on 3-D (multi-band) sources #3100. The source fix stays with that issue; this PR is test-only.
• Backend note: the streaming path is CPU/numpy-only, so no cupy or dask variants apply.

Test plan:

• pytest xrspatial/tests/test_reproject_streaming_3101.py -> 14 passed, 1 xfailed
• No source changes (the only non-test diff is the sweep state CSV)

[brendancol]

PR Review: Add tests for the streaming reproject fallback path (#3101)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• xrspatial/tests/test_reproject_streaming_3101.py:101-123 -- the threaded vs serial split depends on the budget arithmetic inside _process_tile_batch (max_concurrent >= 2 and len(batch) > 1). If the tile_mem formula in _reproject_streaming changes, both tests could silently run the same branch while parity still passes. A cheap guard: assert the computed concurrency in each test (budget // tile_mem >= 2 for the threaded case, == 1 for the serial case) so a formula change that re-routes the branch is visible.
• xrspatial/tests/test_reproject_streaming_3101.py:140-152 -- when #3100 is fixed, the strict xfail flips and forces a revisit, which is the point. The post-fix assertion only checks ndim/shape[2], though. Compare against the in-memory 3-D reproject() output so the un-xfailed test has real teeth from day one.

Nits (optional improvements)

• The dask.bag distributed branch of _reproject_streaming (xrspatial/reproject/__init__.py:1592-1619) stays uncovered; it needs a live dask.distributed client. The sweep CSV records it as a carried LOW, fine to leave -- just noting that "streaming path covered" excludes that branch.

What looks good

• Parity tests compare against the public reproject() output rather than stored fixtures, so the tile assembly is checked against the contract that matters.
• _parse_max_memory cases cover every suffix plus int, float, bare numeric string, and the None default.
• The 3-D crash is pinned strict with raises=ValueError and links the source issue; source stays untouched.

Checklist

• Algorithm matches reference -- N/A, test-only PR
• Backend consistency -- streaming path is CPU-only by design; no GPU variants apply
• NaN handling -- covered by the nearest-resampling NaN-block test
• Edge cases -- single-tile, multi-tile, serial, 3-D strict xfail
• Dask chunk boundaries -- N/A
• No premature materialization -- N/A (tests)
• Benchmark -- not needed for tests
• README feature matrix -- not applicable
• Docstrings -- module docstring explains why the helpers are called directly

[brendancol]

Follow-up review (after d07dd09)

Both suggestions from the first pass are addressed:

• _max_concurrent mirrors the tile_mem arithmetic from _reproject_streaming / _process_tile_batch, and the threaded/serial tests now assert which batching branch they engage (>= 2 vs == 1), plus a multi-tile shape check on the threaded run. A formula change upstream now fails loudly instead of collapsing both tests onto one branch. Fixed.
• test_3d_source_streams compares against the in-memory 3-D reproject() output, so the assertion is full parity once #3100 lands and the strict xfail comes off. Fixed.

Remaining open item is the nit about the dask.bag distributed branch (xrspatial/reproject/__init__.py:1592-1619), deferred with reason: it needs a live dask.distributed client fixture, and it is recorded as a carried LOW in the sweep state CSV.

File re-ran clean: 14 passed, 1 xfailed. No blockers.

[brendancol]

CI note: the first run's macos-3.14 job died with SIGABRT during test_multi_tile_threaded -- two pool threads entered the @njit(parallel=True) projection kernels at once, and numba's workqueue threading layer aborts on concurrent entry. That is a real bug in the streaming path itself (the production ThreadPoolExecutor does the same thing); filed as #3141.

Commit 6432929 keeps the coverage but avoids the unsupported pattern: the threaded parity test runs with transform_precision=0 (exact pyproj, one Transformer per thread), and the multi-tile NaN test plus the 3-D xfail are pinned to the serial loop so the numba fast path is still exercised, just never concurrently. The windows-3.14 "failure" was fail-fast cancellation; its suite had fully passed.