…tensor-stack#2871)

This commit addresses GitHub issue xtensor-stack#2871 where using 'auto' with xtensor
reducer functions (like amax, sum with keep_dims) in functions causes
crashes or incorrect results in optimized builds.

The issue is that lazy expressions hold references to local variables.
When intermediate results are stored with 'auto', these references
become dangling when the function returns, leading to undefined behavior.

Fixes xtensor-stack#2871

- Only force evaluation of the returned expression; intermediate lazy
  expressions are safe as long as they do not outlive the function
- Use auto for shifted/expVals/sumExp in the doc example and tests
- Remove misleading "Alternatively" sections (eval on intermediates is
  not required; evaluation_strategy::immediate does not fix the issue)
- Fix logSoftmax_eval test variant to drop unnecessary xt::eval wrappers

A lazy reducer (e.g. xt::sum with keep_dims) cannot be safely used as an
operand in a subsequent element-wise expression without first materializing
it. Element-wise lazy expressions (shifted, expVals) are safe as auto, but
the reducer result (sumExp) must be assigned to xt::xtensor or wrapped in
xt::eval before use.

This fixes the CI test failure in issue_2871_intermediate_result_handling.