Use segment-local log-log interpolation for profiles with zero-density tail points

[IvanChernyshov]

Problem

interpolate_density(..., mode="loglog") currently uses log-log interpolation only when all density values are strictly positive:

if mode_value == "loglog" and np.all(rho_arr > 0):
    ...
else:
    out = np.interp(...)

Therefore, one numerical zero anywhere in the far tail switches the entire profile to linear interpolation. In the released pbe0_sfx2c_dyallv4z_h-lr_spherical_v2 data, neutral He, F, and Ne contain underflow zeros only at very large radii, so their chemically relevant positive-density regions are nevertheless interpolated linearly.

The present default also returns fill_value=0 below the first grid point. Consequently, a query at r = 0 returns zero rather than the first tabulated near-nuclear density.

Proposed behavior

Make the interpolation rule local to each interval:

• reject negative density values;
• for a positive-positive interval, interpolate in log(r)–log(rho) space;
• for a positive-zero interval, interpolate linearly to zero;
• for a zero-zero interval, return zero;
• preserve exact tabulated values at grid knots;
• for 0 <= r < r_min, return rho(r_min) by default;
• for r > r_max, retain the documented right-side fill behavior, with zero as the current default;
• keep mode="linear" available with its existing meaning.

Non-finite and negative query radii should raise ValueError.

Acceptance criteria

• A zero in the final profile tail no longer changes interpolation in earlier positive-positive intervals.
• Regression tests cover a synthetic positive-positive-positive-zero profile.
• Regression tests cover the released neutral He, F, and Ne profiles.
• r = 0 returns the first tabulated density rather than zero.
• Exact grid-point queries reproduce the stored values.
• Existing fully positive log-log and explicit linear-mode behavior remain covered.
• The numerical contract is documented in the API and methods documentation.

This should require only a helper/API correction and tests; the released profile data do not need to be recomputed.