Fix stale tangent reuse in repeated JVPs

[chrismicah]

Summary

• Replace existing tangent-map entries when a primitive output receives a new JVP.
• Add a regression for the original repeated mx.jvp shape-corruption repro from [BUG] JVP non-deterministic crashes when called in a loop. #3629.

Why

Issue #3629 still reproduces on current main after the earlier partial JVP fixes in #3633 and #3636. On 602b5359, the original ray-trace-shaped repro crashed in 9/10 trials with errors such as:

[squeeze] Cannot squeeze axis 0 with size 125 which is not equal to 1.
[reshape] Cannot reshape array of size 125 into shape (125,4).

The failure is nondeterministic because an older tangent can remain in the JVP transform map when an output id is encountered again. Assigning the output tangent instead of preserving the first inserted value keeps the map aligned with the current primitive output.

Prior related work

• Fix JVPs of select, arctan2, masked_scatter, and bitwise ops #3633 fixed JVPs for select, arctan2, masked_scatter, and bitwise ops.
• Fix JVPs of power, divmod, and slice_update with partly traced inputs #3636 fixed partly-traced JVPs for power, divmod, and slice_update.

Those fixes cover related symptoms, but the original #3629 repro still fails on current main; this PR fixes a distinct current-main regression in the JVP transform map.

Test plan

CMAKE_BUILD_PARALLEL_LEVEL=8 python setup.py build_ext --inplace
PYTHONPATH=python python -m pytest python/tests/test_autograd.py -q -k 'repeated_jvp_with_reused_array_ids or jvp_through_bitwise_ops' --tb=short
PYTHONPATH=python python -m pytest python/tests/test_autograd.py -q --tb=short
python -m black --check python/tests/test_autograd.py
git diff --check

All passed locally.

Fixes #3629