fix(quantization): round INT8 activations to nearest

[zjn20030811]

Summary

• apply the sign-aware half-step while the scaled activation is still floating point
• convert the complete rounded value to INT8 only afterward
• cover positive, negative, and zero activation bins with a Triton interpreter regression

Problem

The current expression converts the +0.5 or -0.5 correction to INT8 before adding it. Both corrections become zero, so the later store truncates every fractional quantization bin toward zero. For a row whose absolute maximum is 1, values +0.5 and -0.5 consequently map to +63 and -63 instead of the nearest bins +64 and -64.

Keeping the correction in floating point until the final conversion restores symmetric round-to-nearest behavior and reduces activation quantization error without changing the scale calculation or kernel interface.

Verification

• python test_cases/test_triton_int8_quantization.py -v (1 test passed)
• ruff check --config pyproject.toml lightx2v/common/ops/mm/triton_kernels.py test_cases/test_triton_int8_quantization.py
• ruff format --check --config pyproject.toml lightx2v/common/ops/mm/triton_kernels.py test_cases/test_triton_int8_quantization.py
• python -m py_compile lightx2v/common/ops/mm/triton_kernels.py test_cases/test_triton_int8_quantization.py