[executorch][devtools] Reset profiler state in ETDump tests

Pull Request resolved: pytorch#21410

Reinitialize the global profiler state before each `ETDumpGenTest` while initializing the runtime once per suite. This prevents profile blocks from leaking across XCTest cases, exhausting the fixed block pool, and causing assertion failures or process crashes. This fixes nine TestInfra signals. This commit was authored with Claude.
ghstack-source-id: 407467887
@exported-using-ghexport

Differential Revision: [D113810097](https://our.internmc.facebook.com/intern/diff/D113810097/)

Author: SS-JIA

runtime/platform/profiler.cpp

@@ -182,6 +182,15 @@ void profiling_create_block(const char* name) {
 }
 
 void profiler_init(void) {
+  prof_header = (prof_header_t*)((uintptr_t)prof_buf + prof_header_offset);
+  prof_arr = (prof_event_t*)((uintptr_t)prof_buf + prof_events_offset);
+  mem_allocator_arr =
+      (prof_allocator_t*)((uintptr_t)prof_buf + prof_mem_alloc_info_offset);
+  mem_prof_arr =
+      (mem_prof_event_t*)((uintptr_t)prof_buf + prof_mem_alloc_events_offset);
+  num_blocks = 0;
+  prof_stats_dumped = false;
+  profile_state_tls = {-1, 0u};
   profiling_create_block("default");
 }