Fix #966 USAN: implicit signed integer truncation in TCX harm model arithmetic coder

• Related issues: #966 (closed)
• Requested reviewers: @multrus @fotopoulou

Reason why this change is needed

• Fix an Undefined Behaviour bug

Description of the change

• The intermediate result of a multiplication of two 16bit values is now stored in an 32bit variable to prevent the implicit cast. The intermediate value is right-shifted afterwards so the result that is finally obtained (in a 16bit var) is always in 16bit range.

Affected operating points

• Everything that uses TCX with harmonic model, but:
• all is BE (stv and ltv)
• bitstream is not changed

Related to #966 (closed)