Fix silent data corruption in the writer, and the packaging and docs

[rafa-guedes]

Stacked on #3. The base is fix/nodal-corrections, so review only shows
this branch's commits; merge #3 first and GitHub will retarget this to master.

Everything here is IO, packaging or documentation. No prediction maths.

Silent corruption

8548324 -- the oceantide format packs into int16 against a fixed scale
(±20 for amplitudes, 0 to 12000 m for depth) and numpy wraps rather than
saturating, so anything past either end round-tripped as a plausible number of
the wrong sign or magnitude, with nothing logged:

25 m amplitude   ->  -15.0 m
15000 m depth    ->  2999.9 m

Both formats. to_oceantide now raises, naming each offending variable and the
range the format can hold. The upper bound is one quantum below the nominal
maximum because the top quantum encodes to the same integer as the fill value
and would decode back as nan. check_range=False opts out of the extra pass
over the data.

This is a deliberate break. Anything writing out-of-range data was already
producing wrong files and will now fail instead.

6576441 -- and the check immediately earned itself. otis_to_oceantide
divides transport by the depth at U and V nodes; that depth is zero at nodes
bordering land, so where the transport was not also zero the result was ±inf.
The land mask keys off the Z-node depth, so those infinities survived into
the result -- five of them in the netcdf fixture alone -- and into written
files. This is what the divide by zero encountered in divide warnings in the
test output have been reporting all along. No flow passes through a dry node,
so the velocity there is zero. Every previously finite value is bit identical,
checked point by point against the old expression on both fixtures, and h, u
and v now share one mask instead of three slightly different ones.

c663b8f -- predictions inherited the codec encoding of the file the
constituents were read from, so predicting from a zarr store and writing the
result back failed with Expected a BytesBytesCodec. The inherited encoding
also names int16 and the fixed scale the constituents are packed against, and
nothing derived from a packed file should be silently re-packed to match it.

Other fixes

d40b9ef -- 0.7.0 taught the accessor to handle datasets carrying only some of
h, u and v, but to_oceantide still indexed self._obj[["dep"]]
unconditionally and raised KeyError: 'dep' on exactly those datasets.
read_oceantide had the mirror problem.

8835055 -- oceantide on the command line printed "Replace this message by
putting your code into oceantide.cli.main". It has been the published entry
point since 0.1.0. Now a small group over the readers and writers that already
exist:

oceantide convert -r otis_netcdf DATA/Model_tpxo9 tpxo9.zarr
oceantide info tpxo9.zarr

8c13d5a -- removes oceantide/ellipse.py, 641 lines of vendored tidal
ellipse code that nothing imports, that has not been runnable since matplotlib
3.0 (plt.hold, plus an exec that assigns to a name read out of a scope it
was never bound in), and that carries its own restrictive licence inside an MIT
package.

60e9d1e -- three packaging problems that all pointed the same way, nothing
was checking:

• pyyaml is imported on every set_attributes call but was never declared;
  it has always been present because dask pulls it in.
• requires-python said >=3.8 and advertised 3.8 to 3.11. The current xarray
  needs 3.11, zarr 3 and numpy need 3.12. Floor is now 3.11, and CI checks it.
• Tests ran only on release. They now run on every push and PR across 3.11 to
  3.13, plus a job pinning zarr<3 -- not padding, the zarr writer branches on
  the major version and picks a different codec import for each, and only the
  installed one was ever exercised.
• The release workflow ran pip install . 'oceantide[extra,test]', which
  installs the working tree and then installs the released version over the
  top, so the tests gating each release ran against the previous release.

d5a2741 -- the packages exclude listed "docs" and "tests", matching those
names exactly and nothing beneath them. Nothing was nested until this branch
added tests/reference, at which point generate.py started shipping in the
wheel.

Documentation

35801ea -- the README was six words under a title and usage.rst said, in
full, "To use oceantide in a project: import oceantide". Both had been the
cookiecutter defaults since 0.1.0, which is plausibly the single biggest reason
nobody outside Oceanum has picked this up.

The README now covers what it does, quick start, formats, constituents and
phase conventions, and a Scope section saying plainly what it does not do and
pointing at pyTMD, eo-tides and UTide instead. usage.rst covers reading,
predicting, sites, staying lazy, amplitude and phase, writing and the CLI.

Every snippet in both was executed against the fixtures before committing --
one of them failed, which is how c663b8f was found.

The docs also could not build at all as configured: conf.py imported
sphinx_rtd_theme while the docs extra installs pydata-sphinx-theme, and
pulled in sphinx_gallery, IPython and matplotlib extensions the extra does not
install either. The gallery those served has been an empty placeholder for six
years and is gone, along with the build artifacts committed under
auto_gallery. sphinx -b html now succeeds.

Verification

113 tests pass, up from 63. twine check passes on both artifacts, the wheel
carries no test or doc files, and a clean install of the built wheel imports,
predicts and runs the CLI outside the source tree.

__version__ is left at 0.8.0; HISTORY.rst has a 0.9.0 (unreleased)
section and the bump is a release decision.

https://claude.ai/code/session_01UwGUDvDZTQWf4MrCig3U1w