Make matplotlib an optional [plot] extra (#2494)

[brendancol]

Closes #2494

matplotlib was listed in install_requires, but nothing in the package imports
it at module top level. Every use is a lazy import inside the .xrs.plot
accessor helpers. So a plain pip install xarray-spatial was dragging in
matplotlib plus its whole chain (pillow, fonttools, kiwisolver, contourpy,
cycler, pyparsing) for users who only run the compute functions.

• Move matplotlib out of install_requires into a new plot extra.
• Add a _require_matplotlib() guard at the two plot entry points in
  accessor.py. Without matplotlib, .xrs.plot() now raises a clear
  ImportError pointing at pip install xarray-spatial[plot] instead of a bare
  ModuleNotFoundError.
• Add matplotlib to the tests extra so the existing plot tests keep running
  in CI.
• Update README, install docs, and CHANGELOG.

pandas is intentionally left alone: xarray requires it and several core modules
import it at top level, so it can't move to an extra.

This is a compute-only change, not backend-specific. No spatial operation
imports matplotlib on any backend.

Test plan

• pytest xrspatial/tests/test_accessor.py (29 passed) — includes new tests
  that block matplotlib in sys.modules and assert core modules still import
  and the plot helpers raise the friendly error.
• pytest xrspatial/geotiff/tests/test_features.py -k "plot or palette or colormap" (8 passed)
• Bare-import check: import xrspatial and the focal/zonal/dasymetric
  modules import with matplotlib blocked.

[brendancol]

PR Review: Make matplotlib an optional [plot] extra (#2494)

Scope is packaging + two guard calls, so the geospatial-specific checks
(backends, dask, CUDA, numerics) don't apply here. I focused on whether the
dependency move is complete and the guard behaves.

Blockers

None.

Suggestions

• _require_matplotlib() (accessor.py:24) checks import matplotlib but the
  call sites then do import matplotlib.pyplot as plt. matplotlib can import
  while pyplot fails (no usable backend, broken install). That's an uncommon
  case and the error would still be an ImportError, just not the friendly one.
  Checking import matplotlib.pyplot in the guard would close the gap. Not
  blocking.

Nits

• test_compute_imports_without_matplotlib calls importlib.reload on
  xrspatial.focal, zonal, and dasymetric. The reload re-executes those
  modules in place, so other tests in the same session keep the reloaded
  objects. It passes today, but reloading numba/ngjit modules for an
  import-side-effect check is heavier than needed. Importing the modules in a
  subprocess with matplotlib blocked, or just asserting the import doesn't
  raise, would isolate the check better.

What looks good

• The core finding holds: no compute path imports matplotlib. The only runtime
  uses are the two .xrs.plot accessor methods; the matches in zonal.py and
  bump.py are inside .. plot:: docstring examples, not executed code.
• matplotlib added to the tests extra, so the existing plot/palette tests keep
  running in CI rather than silently skipping.
• pandas correctly left as required, with the reason recorded.
• _listed_colormap_from_attrs kept as-is; it already returns None on
  ImportError and sits on the read path, so guarding it would have broken
  non-plotting GeoTIFF reads.
• New tests assert both the friendly-error path and that core modules import
  with matplotlib blocked.

Checklist

• No compute path imports matplotlib (verified across the package)
• Friendly ImportError points at the plot extra
• tests extra keeps plot tests running in CI
• Docs (README, installation.rst) and CHANGELOG updated
• pandas correctly identified as non-removable
• Guard covers matplotlib.pyplot, not just matplotlib (see suggestion)

[brendancol]

Follow-up review (after 1e1157d)

Both findings from the first pass are resolved.

• Suggestion (guard pyplot): _require_matplotlib() now does
  import matplotlib.pyplot, so an installed-but-unusable matplotlib also
  produces the friendly error. Resolved.
• Nit (reload-based test): test_compute_imports_without_matplotlib now runs
  in a subprocess against a clean module cache instead of reloading numba
  modules in-session. Resolved.

No new issues. pytest xrspatial/tests/test_accessor.py is 29 passed and
flake8 is clean on the changed files. No remaining items beyond CI.