Since 7ab4c928, osdep.h contains
calls to malloc/free.

This fixes building with MSVC targeting WinRT.

Approach taken from x86 asm. Overall speedup meaningless.
cabac_encode_terminal on average twice as fast on cortex-53 while
encoding with following command:
./x264 --threads 1 --profile high --preset veryfast --crf 15 -o /dev/null park_joy_420_720p50.y4m

Less relative speedup on cortex-a72/73.

Imported from dav1d.

0.5% - 2% overall speedup on
`./x264 --threads X --profile high --preset veryfast --crf 15 -o /dev/null park_joy_420_720p50.y4m`
cabac is responsible for roughly 1/6 of the CPU use.
Branch mispredictions are reduced by 15% to 20%.

cortex-s53: 0.5% faster
cortex-a72: 2%  faster
neoverse-n1: 0.9% faster

It became unused in cde9a933.

Anton Mitrofanov authored Jan 22, 2021 and Anton Mitrofanov committed Jan 24, 2021

This bug mainly occurred when encoding with high bitrate (low QP).
It did not occur when encoding in baseline or main profile.

Anton Mitrofanov authored Jan 20, 2021 and Anton Mitrofanov committed Jan 24, 2021

This bug never occurs with the current reference management logic.
Bug report by Lingjiang Fang.

Anton Mitrofanov authored Jan 21, 2021 and Anton Mitrofanov committed Jan 24, 2021

Broke after unifying of 8-bit and 10-bit libraries.