feat: plain strings emit DeprecationWarning

[andremsouza]

The warning half of the string deprecation. 0.4.0 gave the written notice; this is what follows it,
and it is the last thing 1.0.0 needs before it can remove strings.

DeprecationWarning: Passing mode='batch' as a plain string is deprecated and will
stop working in 1.0.0. Use TrainingMode.BATCH instead.

The message names the exact replacement. A deprecation that makes the reader work out the
substitution is a deprecation people silence.

Two cases that deliberately stay silent

Both were found by building it, not by planning it.

An invalid string does not warn. My first version warned before validating, so under
filterwarnings = ["error"] a typo produced a DeprecationWarning instead of its ValueError —
burying the real mistake under a notice telling the caller to modernise a spelling that never worked.
Reordering fixed the obvious cases but not sample_mode="cauchy", which is validated further in. The
rule that actually generalises is warn only for values the enum recognises, which removes the
ordering question entirely.

Loading a map does not warn. This one was a genuine design bug. load_npz reconstructs a map by
passing the stored strategy name to the constructor — and that name is a string, because JSON has
no enums. So opening any saved file warned the user about a string the file chose and they never
wrote. The loader now converts it to the member it denotes. mexicanhat has no member of its own
and is an alias for the same function, so it maps to MEXICAN_HAT.

A regression this surfaced

Error messages interpolated the value directly, so passing an enum produced:

<WeightInit.LINEAR: 'linear'> initialization requires a 'data' argument

Both spellings now produce identical text, asserted by a test that compares the two.

The suite

filterwarnings = ["error"] means every string call site had to be dealt with rather than ignored.
55 sites moved to enum members across 8 files. The tests that exist to prove both spellings work
now assert the warning too, since that is the behaviour under test:

error_enum = first.train(blobs, n_iteration=15, mode=as_enum)
with pytest.warns(DeprecationWarning, match=f"mode={as_string!r}"):
    error_string = second.train(blobs, n_iteration=15, mode=as_string)
assert error_enum == error_string

Dedicated tests cover the deprecation itself: the message names the replacement for every spelling
including the legacy one, an invalid spelling raises instead of warning, an enum never warns, the
warning blames the caller's line rather than a line inside the package, and a stored artifact name
converts rather than warning.

Verification

384 tests, 100% coverage, ruff / ruff format / mypy --strict clean, mkdocs build --strict
exits 0, bandit clean, architecture profile --strict passes.

Not in this PR

The version bump. This is the change; the release is its own PR, as with 0.4.0.