[doc] Fix exp() single-precision abs errors.

[wjakob]

I think that's actually correct. Even a 1 ULP error can translate into a huge absolute error when working with the exponential function due to its large output.

[SepandKashani]

After some more digging, I believe you are right: the abs/rel errors obtained numerically are close to those listed in the tables.

from mpmath import mp
import drjit as dr
import drjit.llvm as drb

N = 1_000
a32 = dr.linspace(drb.Float, -20, 30, N)
a64 = dr.linspace(drb.Float64, -20, 30, N)
b32 = dr.exp(a32)
b64 = dr.exp(a64)

with mp.workdps(8):
    gt32 = list(map(mp.exp, a32))
with mp.workdps(16):
    gt64 = list(map(mp.exp, a64))

abs32, rel32 = [], []
abs64, rel64 = [], []
for _b, _gt in zip(b32, gt32):
    abs32.append(abs(_b - _gt))
    rel32.append(abs(_b - _gt) / _gt)
for _b, _gt in zip(b64, gt64):
    abs64.append(abs(_b - _gt))
    rel64.append(abs(_b - _gt) / _gt)

max(abs32), max(rel32), max(abs64), max(rel64)

# (mpf('770048.0'),
#  mpf('1.078238811961534729e-7'),
#  mpf('0.000732421875'),
#  mpf('2.52219106387520338e-16'))

Sorry for the spam; this PR can be closed.