fix: withdraw the plain-string deprecation

[andremsouza]

0.5.0 deprecated the plain-string options and announced removal in 1.0.0. That was wrong, and this
withdraws it. Strings are permanently supported.

Why it was wrong

Every comparable library passes options as plain strings, and none export enums at all. Measured
against the installed versions, not recalled:

library	how options are passed	enums exported
scikit-learn 1.9	KMeans(init="k-means++", algorithm="lloyd")	none
numpy 2.5	np.pad(mode="constant"), norm(ord="fro")	none public
scipy 1.18	linkage(method="single"), minimize(method="BFGS")	none
minisom	neighborhood_function="gaussian", topology="rectangular"	none
sompy	normalization="var", neighborhood="gaussian"	none

scikit-learn validates them at runtime with StrOptions — a constraint, not a type. minisom is the
closest peer, and python-som's method names already mirror it (activate, winner, quantization,
quantization_error, get_weights).

Removing the string form would have made this the only library in its ecosystem to reject
mode="batch".

The book said less than I claimed

The original design cited Viafore, Robust Python, ch. 8. Re-read, his case is about preventing typos
inside a codebase, and he includes a "When Not to Use" section. The benefit he actually claims is
that a type checker rejects a wrong value — and the Literal unions added in 0.4.0 already deliver
exactly that, with strings still working:

som.train(data, mode="batch")             # fine
som.train(data, mode=TrainingMode.BATCH)  # fine
som.train(data, mode="bacth")             # error: incompatible type Literal['bacth']

So the deprecation bought nothing that was not already had, and cost ecosystem divergence.

What changes

• warn_if_string and its four call sites are removed.
• The tests that wrapped string calls in pytest.warns are unwrapped.
• A new test asserts the absence of any warning at every call site, under
  warnings.simplefilter("error") rather than the removed pytest.warns(None), so any warning at all
  fails. The withdrawal cannot be quietly undone.
• The docs section becomes "Both spellings are permanent", with the survey table above.

The enums stay. They cost nothing, some callers prefer them, and removing them would be a second
reversal in two releases. Anyone who migrated while 0.5.0 was current keeps working code.

Where this leaves the roadmap

refactor/enums-only is dropped from 1.0.0. With strings staying, what remains breaking is thin —
keyword-only arguments, removing two setters, removing pre-0.3.0 private aliases, dropping Python 3.10
after its October 2026 EOL — and the plan already warned against freezing an interface before it has
been used. So 0.6.0 adds the estimator API additively and 1.0.0 waits for real use of it.

On process

The deprecation was decided without checking what comparable libraries do. Reversing one release later
is cheap; discovering it at 1.0.0, after breaking every user, would not have been.

Verification

383 tests, 100% coverage, ruff / ruff format / mypy --strict clean, mkdocs build --strict exits
0, architecture profile --strict passes. Every call site exercised with plain strings under
-W error::DeprecationWarning.