[k2] use initial-exec TLS model for performance critical thread locals

[apolyakov]

The initial-exec TLS model provides faster access to thread-local variables by using a more efficient addressing mode that avoids dynamic TLS resolution at runtime. Since these state pointers are accessed frequently in hot paths, caching them in thread-local storage with initial-exec model improves performance.

The state pointers are now set once during lifecycle transitions (image/component/instance creation and initialization, and at poll entry) and then accessed directly from TLS rather than calling through the C API layer each time.

Note

The initial-exec TLS model is generally unsafe for dynamically loaded libraries (it requires TLS to be allocated at program startup). This is acceptable for K2 because the runtime is loaded in a controlled environment where this constraint is satisfied. In case of any problems we can try -mtls-dialect=gnu2 which is safe but a bit slower than initial-exec TLS model.