Fix shapeless compile eliding reductions over size-1 dimensions

[discobot]

Proposed changes

Fixes #3201.

This picks up where #3202 left off: reductions over dimensions that are size 1
at trace time are elided from the graph, so shapeless replays with larger
dynamic shapes return stale values.

The reason #3202 failed CI was not a numerical regression. Once the reduction
is kept in the graph it can be a runtime identity (out.size() == in.size(),
e.g. a replay with size 1), and Reduce::eval_gpu rejects that case with
debug-only asserts in both the Metal and CUDA backends — which is why the
CPU-only jobs passed while the macOS and CUDA jobs aborted on the PR's own new
test. This PR replaces those asserts with an explicit identity path that
cast-copies the input to the output (the CPU backend already computes the
identity case correctly).

The frontend change is also restricted to non-empty reduction axes. The
unconditional version in #3202 could create a Reduce with no axes (0-dim
input, or an empty axes list for all/any/max), tripping
assert(!axes_.empty()) in the same functions. No-axis reductions are shape
independent so eliding them stays correct.

The regression test covers all eight reductions plus mean on the output of
mx.take, in both directions: trace at size 1 and replay larger (the original
bug), and trace at size 2 and replay at size 1 (the runtime-identity path that
broke #3202). Verified against a Debug (assertions enabled) CPU build;
test_compile, test_reduce, and test_ops pass.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

Notes on the pre-commit box: clang-format (the three C++ files) and
black/isort --profile=black (the test file) were run and report clean, but
with clang-format v22.1.5 rather than the v21.1.8 the repo pins, and the full
pre-commit run --all-files was not run, so I have left that box unchecked for
the maintainer to confirm.

This change was developed with the help of Claude Code.