fix: [libpthreadglue] tighten error handling in OS-abstraction layer

[fjtrujy]

Summary

• pte_osThreadCreate ignored the return value of sceKernelCreateSema when creating cancelSem, and on the sceKernelCreateThread failure paths it leaked the just-created cancelSem.
• Reorder so each resource is validated immediately after it is allocated and cleaned up in reverse order on failure.
• Propagate sceKernelCreateSema failures from pte_osMutexCreate and pte_osSemaphoreCreate as PTE_OS_NO_RESOURCES instead of writing a negative handle to *pHandle as if it were a valid semaphore.

Related

• Companion pthread-embedded PR: Reject NULL / uninitialised mutex handles with EINVAL pthread-embedded#3 — adds the NULL-handle rejection and mutex9 test referenced in the test plan.

Test plan

• Build src/libpthreadglue cleanly.
• Run pthread-embedded test suite on PSP — exercises thread/mutex/sema creation paths and now also the NULL-handle paths from the companion pthread-embedded PR.
• Smoke-test a sample that creates pthreads.