fix: dtype handling and axis validation in statistical functions

[henryiii]

🤖 AI text below 🤖

Fixes three confirmed bugs in src/ragged/_spec_statistical_functions.py.

1. sum / prod ignored their dtype argument

The cast condition was inverted: it cast only when the dtypes already
matched (a no-op) and skipped the cast when they actually differed.

• Before: ragged.sum(ragged.array([[1, 2, 3], [4, 5]]), dtype=np.float32) returned int64.
• After: returns float32 (value 15.0). The cast now happens via the
  existing _ensure_dtype helper, which also handles the 0-d scalar case.

2. mean / var / std truncated integer input

These functions cast their floating-point result back to x.dtype,
truncating integer input.

• Before: ragged.mean(ragged.array([1, 2])) returned 1;
  ragged.var(ragged.array([1, 2, 4])) returned 1 (true ≈ 1.556).
• After: integer/bool input keeps its natural floating-point dtype
  (mean([1, 2]) == 1.5), matching the Array API requirement that these
  functions return a floating-point data type. Floating input (e.g.
  float32) is still cast back to preserve the "same data type as x"
  promise. Docstrings' "Returns" wording updated accordingly. std
  delegates to var, so it follows automatically.

3. _regularize_axis built an error message but never raised it

For tuple axes, the out-of-bounds branch assigned msg but never raised,
and the bound check 0 < out[-1] incorrectly rejected axis 0.

• Before: ragged.sum(x, axis=(0, 5)) fell through to a less helpful
  awkward error.
• After: raises ak.errors.AxisError with a clear message; the bound
  is now 0 <= out[-1] < ndim, so tuple axes including 0 work.

Tests

Added regression tests in tests/test_spec_statistical_functions.py
covering all three fixes. Full suite passes (1305 passed, 6 skipped),
mypy --strict and all pre-commit hooks clean.

Part of #144

🤖 Generated with Claude Code