Problem

HeatMap's docstring documents "a numpy.array of shape (n,2) or (n,3)" as
valid input, but an all-integer array crashes at render time:

import numpy as np, folium
from folium.plugins import HeatMap

m = folium.Map()
HeatMap(np.array([[3, 4, 1], [5, 6, 2]])).add_to(m)   # int64 dtype
m.get_root().render()
# TypeError: Object of type int64 is not JSON serializable

The float-weight version renders fine. This is common in real usage, e.g.
HeatMap(df[["lat", "lon", "count"]].values) where the weight column has an
integer dtype (pandas .values yields an int64 array).

Cause

In HeatMap.__init__, validate_location coerces the lat/lon columns to Python
float, but the weight column (line[2:]) is passed through unchanged:

self.data = [
    [*validate_location(line[:2]), *line[2:]] for line in data
]

np.float64 is a subclass of float, so json.dumps (via the template's
|tojson) tolerates float weights; np.int64 is not a subclass of int,
so integer weights raise. The existing test_heatmap_data masks this because
its array contains a 0.5, forcing the whole array to float64.

Fix

Coerce the weight column to float, mirroring the lat/lon normalization. The
NaN guard immediately below still works (float(np.nan) is caught by
np.isnan).

Tests

• test_heatmap_integer_numpy_weights: an int64 (n,3) array — weights
  normalized to Python float, render contains L.heatLayer, and integer vs
  float arrays produce identical .data.
• test_heatmap_integer_numpy_no_weight: an int64 (n,2) array renders.

pytest tests/plugins/ passes (the one unrelated test_time_slider_choropleth
failure is a pre-existing missing optional dependency, identical on a clean
checkout). ruff check clean.