gh-40928: Avoid swallowing AlarmInterrupt in matrix destructor

[cxzhong]

Fixes #40928.

Root cause

linear_code.py deliberately interrupts a costly canonical-representative
computation after 0.5 seconds. When the alarm arrived while a temporary FLINT
matrix was being destroyed, Matrix_modn_dense_flint.__dealloc__ entered a
sig_on() region. This raised AlarmInterrupt from the destructor, where the
exception could not propagate and was reported as ignored. Because the alarm
was one-shot, the computation then continued until the doctest file timeout.

This follows #18087, which established that __dealloc__ methods cannot raise
exceptions and therefore must not enter sig_on() regions.

Changes

• Do not wrap nmod_mat_clear() in sig_on()/sig_off() during destruction.
• Initialize _shift_mod's preshift temporary explicitly, eliminating the
  compiler's -Wmaybe-uninitialized warning.

Validation

• Rebuilt matrix_modn_dense_flint successfully; the compiler warning is gone.
• src/sage/matrix/matrix_modn_dense_flint.pyx: 175 doctests passed.
• src/sage/coding/linear_code.py: 424 doctests passed.
• Repeated the linear-code doctests with all four random seeds from recent CI
  timeouts; all passed.

Local validation used Linux with Python 3.12; CI will additionally exercise the
macOS/Python 3.14 configuration where this occurred frequently.

[github-actions]

Documentation preview for this PR (built with commit fe1e853; changes) is ready! 🎉
This preview will update shortly after each push to this PR.

[vincentmacri]

Test failure is real but is unrelated as it occurs on develop. Reported in #42623.