A few correctness fixes

[astrofrog]

This fixes a few bugs:

• integral_general_subset adds spurious area outside the tabulated range. When the integration limits extend beyond the tabulated x, the routine adds a triangular ramp of area past the domain instead of clamping. Fix: clamp the effective limits to [x(1), x(n)] so out-of-range extremities contribute zero.

• ipos mishandles both boundaries for descending ranges (xmax < xmin). x == xmin and x == xmax return off-by-one bins. Fix: correct the boundary handling so xmin maps to bin 1 and xmax to bin nbin.

• xval computes bin centres in single precision. The bin-centre fraction (i - 0.5) / nbin is evaluated in default (single) precision before being used, so the double-precision instantiation loses precision. Fix: compute the fraction in the instantiated kind.

• sample_pdf_log returns NaN for draws in the first CDF bin. A random value landing in the first bin feeds cdf(1) = 0 into log-log interpolation (log(0)), producing NaN. Fix: use linear interpolation on the CDF axis with boundary guards.