Make aten::logaddexp, logaddexp2 and log_sigmoid numerically stable

[Raashish Aggarwal (raashish1601)]

aten_logaddexp, aten_logaddexp2 and aten_log_sigmoid used the direct formulas, which overflow or underflow well inside the float32 range:

• logaddexp(a, b) = log(exp(a) + exp(b)): for a, b around 100 or more exp overflows and the result is inf (PyTorch returns e.g. 1000.69 for logaddexp(1000, 1000)).
• logaddexp2 has the same problem with 2**a + 2**b.
• log_sigmoid(x) = log(sigmoid(x)): for x below about -100, sigmoid underflows to 0 and the result is -inf instead of about x.

They now use the usual stable forms (the ones PyTorch uses):

• logaddexp: max(a, b) + log(1 + exp(-|a - b|))
• logaddexp2: max(a, b) + log2(1 + 2**-|a - b|)
• log_sigmoid: min(x, 0) - log(1 + exp(-|x|))

For a == b == +/-inf, a - b is NaN, so logaddexp/logaddexp2 return a + log(2) / a + 1 when the inputs are equal, which gives the right infinity (and the same value as the formula for finite inputs).

Testing:

• Added ops.aten.logaddexp.large, ops.aten.logaddexp2.large and ops.aten.log_sigmoid.large OpInfos with values from -1000 to 1000 and +/-inf (all pairs for the binary ops). All three fail on main and pass with the change.
• pytest tests/function_libs/torch_lib/ops_test.py -k "logaddexp or log_sigmoid or logsigmoid" passes locally, and an exported model with these ops matches PyTorch on large inputs through onnxruntime. ruff check / format are clean on the changed files.