Fix integer overflow in single rounding requantization

[veblush]

Problem

When CMSIS_NN_USE_SINGLE_ROUNDING is enabled, arm_nn_requantize suffered from an integer overflow. The intermediate value evaluated as new_val >> (total_shift - 1) can require up to 33 bits. Truncating this into an int32_t prematurely flipped the sign bit for large positive values, resulting in catastrophic output errors (e.g., yielding heavily negative numbers instead of saturated maximums).

Solution

• Promoted the intermediate result variable from int32_t to int64_t to safely hold the 33-bit value.
• Moved the int32_t cast to the final return after the >> 1 shift has brought the value safely back into 32-bit bounds.

[mansnils]

Thanks for this Esun!

Could you also update the date/revision? New version v22.9.1.