Fix int32_t-vs-int portability; align CThread assignment with std::thread

[aklofas]

Summary

Three small portability fixes that let libsrt compile and run on toolchains where int32_t is a distinct type from int (e.g. the newlib arm-none-eabi bare-metal toolchain, where int32_t is long int) and on pthread backends without thread cancellation.

Reworked per review (see comments): the config field is typed to match the wire, TEV_SYNC passes no argument, and CThread assignment matches std::thread — no opt-out define needed anymore.

Changes

1. CSrtConfig::iMSS is now int32_t
Matching the handshake wire field CHandShake::m_iMSS it is clamped against, so std::min(m_config.iMSS, w_hs.m_iMSS) in acceptAndRespond() deduces its template argument. The SRTO_UDP_SNDBUF/SRTO_UDP_RCVBUF setters that std::max the option value against iMSS read it with cast_optval<int32_t> for the same reason. Where int32_t == int this is a pure no-op.

2. common.h/core.h — EventVariant() no-argument constructor
EventVariant(0) at the TEV_SYNC call site was ambiguous between EventVariant(int32_t) and EventVariant(const CPacket*) once int32_t is not int. TEV_SYNC passes no argument (no slot is connected to it, and updateCC never reads the arg for it), so EventVariant now has a no-argument constructor producing UNDEFINED, used at that call site.

3. sync_posix.cpp — CThread assign-to-joinable matches std::thread
CThread::operator= called pthread_cancel() on a still-joinable thread — guarded out on Android/OHOS (#1476 / #1484) and missing on some RTOS pthread backends (e.g. FreeRTOS-Plus-POSIX). Per std::thread semantics ([thread.thread.assign]) and the ENABLE_STDCXX_SYNC backend, this is now std::terminate() after the existing IPE log line. This removes the only pthread_cancel call in the codebase, so no platform special-casing is needed at all.

Testing

Built and run on bare-metal arm-none-eabi (Cortex-M4F) under QEMU on a FreeRTOS + lwIP substrate (pthread sync backend, ENABLE_STDCXX_SYNC=OFF): libsrt cross-compiles cleanly; an srt_startup → srt_create_socket → srt_getsockstate → srt_close → srt_cleanup smoke test and a caller↔listener loopback transfer over a lossy link (plain and AES-128 via mbedTLS) pass. Desktop Release build with ENABLE_UNITTESTS=ON compiles unchanged and the full test-srt suite passes (275/275).

Behavior note: the only behavior change on existing platforms is the assign-to-joinable IPE path, which now terminates — as std::thread and the ENABLE_STDCXX_SYNC build already do — instead of cancelling the live thread.

[ethouris]

I'd have different changes in mind.

1. For the config.iMSS field, this type should be changed into int32_t.
2. For EventVariant, that case with TEV_SYNC is that it gets no argument - so this is simply a sealup replacement. A better solution would be to provide the version with no arguments, which results in UNDEFINED type inside. The argument isn't used anyway.
3. The CThread should behave exactly the same way as std::thread, as far as possible. Not sure why it was done this way, but I think CThread version for POSIX should do the same as std::thread - that is, if (joinable()) std::terminate();. That should also solve the problem of missing pthread_cancel everywhere if we stop using it.

[aklofas]

Reworked along the lines you suggested — all three points, and the branch is updated:

1. CSrtConfig::iMSS is now int32_t. One ripple to be aware of: the SRTO_UDP_SNDBUF/SRTO_UDP_RCVBUF setters std::max the option value against iMSS, so they now read it with cast_optval<int32_t> — otherwise they'd hit the same deduction failure on platforms where int32_t ≠ int. With the field typed correctly, the std::min<int> pin is reverted; core.cpp is no longer touched by this PR.

2. EventVariant has a no-argument constructor producing UNDEFINED, and the TEV_SYNC call site uses it. I checked that nothing reads the argument — no slot is ever connected to TEV_SYNC, and the generic TEV_SYNC handling in updateCC doesn't touch arg.

3. POSIX CThread::operator= now calls std::terminate() when assigning to a joinable thread, matching std::thread ([thread.thread.assign]) and the ENABLE_STDCXX_SYNC backend. That removes the only pthread_cancel() call in the codebase, so the SRT_NO_PTHREAD_CANCEL define I proposed is gone — and the existing __ANDROID__/__OHOS__ special case ([BUG]build for Android can't use pthread_cancel #1476) with it. I kept the IPE log line ahead of the terminate so the abort is diagnosable.

Re-verified on both ends: desktop Release build with the full test-srt suite green (275/275), and the original bare-metal target (arm-none-eabi Cortex-M4F, FreeRTOS + lwIP under QEMU, where int32_t is long int) — cross-compiles cleanly and passes an srt_startup/socket smoke test plus a caller↔listener loopback transfer over a lossy link, plain and AES-128. I've updated the PR title/description to match.