Fix/cummax cummin pyarrow float

[genrichez]

BUG: fix cummax/cummin for pyarrow float dtypes

Closes #66257

What's wrong

Series.cummax() on pyarrow float columns returns garbage when the running max is negative. For example:

s = pd.Series([-4.0, -3.0, -2.0, -1.0], dtype="float64[pyarrow]")
s.cummax()
# returns [2.2e-308, 2.2e-308, 2.2e-308, 2.2e-308] instead of [-4, -3, -2, -1]

Same idea for cummin when the running min is +inf.

Why it happens

pyarrow.compute.cumulative_max defaults its start parameter to the smallest positive normal float (~2.2e-308 for float64), not -inf. So any negative value loses to that default. Similarly cumulative_min defaults to the largest finite float, not +inf.

The fix

Pass start=-inf for cummax and start=+inf for cummin when the array dtype is floating. This matches numpy behavior exactly. Integer and temporal types are unaffected (pyarrow already uses correct defaults for those).

if pa.types.is_floating(pa_dtype) and name in ("cummax", "cummin"):
    start_value = float("-inf") if name == "cummax" else float("inf")
    kwargs.setdefault("start", pa.scalar(start_value, type=pa_dtype))

Tests

Added regression tests covering:

• All negative values (the reported bug)
• +inf in cummin
• Mixed positive/negative matching numpy output
• NA handling with skipna True/False
• Both float32 and float64