Narrow float64 histogram ranges mis-bin pixels (halved range rounds to zero, n / half_range overflows) #84

From an external static review (batch 18, N4), reproduced on main at 567c20b.

Where: histogram.py:665-680 (_float_bins), histogram.py:851-864 (_bin_index), bin_centres.py:16.

• half_range = hi*0.5 - lo*0.5 can round to zero for distinct subnormal endpoints. float_bin_centres then collapses distinct bins.
• n / half_range can overflow to inf for a very narrow non-zero range, before the integer bin cast.
• Either way, finite pixels can be mis-binned and pooled percentiles corrupted.

Fix: compute a bounded ratio first and multiply by n after. When hi - lo is finite, use (value - lo) / (hi - lo). Keep the halved form only for the large-range case where hi - lo overflows. Apply the same rule to bin centres and _bin_index. The large-range fix of #76 (±float64 max) must still pass.

Reproduction on main:

• float64 plane [0, 5e-324, ...], 4 bins: ZeroDivisionError (range check path / _bin_index).
• plane [0, 1e-320, 0, 1e-320], 4 bins: counts [4, 0, 0, 0] (the 1e-320 pixels land in bin 0) plus RuntimeWarning: invalid value encountered in cast.
• plane [0, 5e-307, 1e-306], 1024 bins: n / half_range is inf, so the position is nan or inf before the cast.

Regressions wanted: the smallest-subnormal range and a narrow normal range; existing histograms must stay bit-identical.