Fix 32-bit regressions in ndarray encode/decode and stochastic rounding tests

[awf]

Title

Fix 32-bit regressions in ndarray encode/decode and stochastic rounding tests

Description

Summary

This PR resolves issue #57 by removing temporary 32-bit xfails and fixing the underlying dtype/bit-packing issues that caused i386 failures.

Changes

• Fixes ndarray decode path to use 32-bit-safe ldexp exponent/input dtypes.
• Fixes ndarray encode path bitfield packing to avoid sign-extension/overflow behavior on 32-bit.
• Updates stochastic-rounding test random-bit generation to use explicit int64 so 2**32 bounds work on 32-bit Python.
• Restores intended stochastic test coverage (srnumbits=32) after dtype-safe generation.
• Updates linux/386 test script argument handling so:
  • trailing args are forwarded to pytest
  • default test target is only used when no args are supplied
  • sh ... invocation remains robust

Validation

• sh etc/test-linux-386.sh run -vv test/test_encode.py::test_encode[binary32] ✅
• sh etc/test-linux-386.sh run ✅ (913 passed, 3 skipped)
• Local focused suite on modified areas ✅ (872 passed)

Context

Closes #57.