fix(registry): bump docstring truncation cap to 4000 chars (#335)

[SAY-5]

Closes #335.

What

EstimatorNode.to_dict truncated the estimator docstring at 500
characters before exposing it to the LLM. sktime's numpydoc-style
docstrings put the Parameters section, the most useful piece for
an agent reasoning about hyperparameters, past the first 500 chars
in many estimators (composite forecasters, classifiers with 10+
parameters, etc.), so the truncation hid exactly the content the LLM
needed to configure the model.

Fix

Bump the cap to 4000 characters via a new module-level constant
(_DOCSTRING_MAX_CHARS) and a helper (_truncate_docstring):

• short / None docstrings pass through unchanged
• longer ones get truncated to the cap with a trailing ... marker
  so the consumer knows content was elided
• 4000 is large enough to keep the Parameters block of every
  sktime estimator I sampled while still keeping the MCP tool result
  reasonably compact

The issue suggested either bumping the limit or implementing a
"prioritize Parameters section" parser. I went with the simpler cap
bump since smarter parsing has a much larger surface (numpydoc /
Google-style / plain text variations); happy to follow up with a
section-aware truncator if you'd prefer that direction.

Tests

tests/test_core.py::TestDocstringTruncation:

• test_short_docstring_unchanged
• test_none_passes_through
• test_long_docstring_truncated_with_ellipsis
• test_cap_preserves_parameters_section_past_500_chars, the
  [BUG] Hardcoded 500-character docstring truncation limits LLM reasoning #335 regression: docstring with a Parameters block starting at
  offset 600 must still surface after truncation.

Verification

• pytest tests/test_core.py::TestDocstringTruncation -v → 4 passed.
• Reverting _DOCSTRING_MAX_CHARS to 500 makes the last subtest
  fail with assert 'Parameters' in '<x...>', confirming the test
  catches the regression.