- fix "здж" > "ждж" rule, fix test for it
- add rules for "б г д дз дж з" at the end of words
- added tests for new rules
- add devoiced rules conversion and tests
- languages.md: fix the Belarusian language name

fix "здж" > "ждж" rule, fix test for it
add rules for "б г д дз дж з" at the end of words
add devoiced rules conversion
add some words to vocabulary
added tests for new rules

This reverts commit 8021528f5e.

Failing on CI/CD.

Code cleanups found with static analyzers.

StatusCreateTerminatedMsg.

This is not covered by current tests.

Checking for return value & continue might be redundant.

Found with pmd-cpd static analysis

This fixes the intonation in Latin American Spanish questions, removes the phoneme replacement (T to s) because it is already included in phonemes es_la. It also eliminates the line languaje es 6, because when the voice is in the dialect of Spain and it automatically changes to the dialect of Latin America, it does not return to the dialect of Spain.

When compile from read-only data bundle (shipped within iOS/macOS app),
we can't write compile_report file.
The simplest solution is to (almost) silently discard it.

This fixes the intonation in Latin American Spanish questions, removes
the phoneme replacement (T to s) because it is already included in
phonemes es_la. It also eliminates the line languaje es 6, because when
the voice is in the dialect of Spain and it automatically changes to the
dialect of Latin America, it does not return to the dialect of Spain.

Note: this doesn't fix actual problem for some of the older devices.
Unfortunately i don't have devices with os version below than android 7,
so am not able to test.

On e.g. iOS it may be very long.

Helps to make #1291

So we can ship data bundle and compile it AOT.

Helps to make #1291

Fixes for some code smells found with cppcheck static code analysis tool.

Strlen never retuns a negative value. Also combined if statements to
reduce complexity. This code currently has no tests.

Inner 'n_bytes >= 0' will always be true.

Inner 'p->stresslevel>=4' will always be true. Thus checking for
count_stressed is redundant.

Unify common parts of two implementations. Split into smaller functions.

Google is now a LLC not Inc.

This allows to cover
- another arch that x86, which is also very common
- an arch where char is unsigned
- a little-endian arch

This is running under qemu-user, so significantly slower and doesn't
support ptrace thus no address sanitizer etc. but at least the undefined
sanitizer works.

implemantations.

It is unclear why the two implementations differ.