adjust header inclusions in C10 as sugguested by IWYU (#102467)

This PR aims to reduce unused header inclusions in C10.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/102467
Approved by: https://github.com/albanD

c10/core/DispatchKey.h

@@ -2,6 +2,7 @@
 
 #include <c10/core/DeviceType.h>
 #include <c10/macros/Export.h>
+#include <cstdint>
 #include <ostream>
 #include <string>
 

c10/core/StorageImpl.cpp

@@ -1,4 +1,5 @@
 #include <c10/core/StorageImpl.h>
+#include <c10/util/flat_hash_map.h>
 
 namespace c10 {
 

c10/core/StorageImpl.h

@@ -4,8 +4,6 @@
 #include <c10/core/SymInt.h>
 #include <c10/core/impl/PyObjectSlot.h>
 
-#include <c10/util/Optional.h>
-#include <c10/util/flat_hash_map.h>
 #include <c10/util/intrusive_ptr.h>
 
 namespace c10 {

c10/core/impl/LocalDispatchKeySet.cpp

@@ -1,7 +1,5 @@
 #include <c10/core/impl/LocalDispatchKeySet.h>
 
-#include <iostream>
-
 namespace c10 {
 namespace impl {
 

c10/core/impl/LocalDispatchKeySet.h

@@ -1,8 +1,7 @@
 #pragma once
 
 #include <c10/core/DispatchKeySet.h>
-#include <c10/macros/Macros.h>
-#include <c10/util/Flags.h>
+#include <c10/macros/Export.h>
 
 // TLS management for DispatchKeySet (the "local" DispatchKeySet(s))
 //

c10/core/impl/cow/context.cpp

@@ -2,6 +2,7 @@
 
 #include <c10/core/impl/cow/deleter.h>
 #include <c10/util/Exception.h>
+#include <mutex>
 
 namespace c10::impl {
 

c10/core/impl/cow/context.h

@@ -1,12 +1,11 @@
 #pragma once
 
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 #include <c10/util/UniqueVoidPtr.h>
 
 #include <atomic>
 #include <cstdint>
 #include <memory>
-#include <mutex>
 #include <shared_mutex>
 #include <variant>
 

c10/core/impl/cow/deleter.h

@@ -5,7 +5,7 @@
 
 #pragma once
 
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 
 namespace c10 {
 namespace impl {

c10/mobile/CPUCachingAllocator.cpp

@@ -1,6 +1,6 @@
-#include <c10/mobile/CPUCachingAllocator.h>
-
 #include <c10/core/impl/alloc_cpu.h>
+#include <c10/mobile/CPUCachingAllocator.h>
+#include <c10/util/Exception.h>
 
 namespace c10 {
 

c10/mobile/CPUCachingAllocator.h

@@ -1,11 +1,7 @@
 #pragma once
 
-#include <algorithm>
-#include <deque>
-#include <memory>
 #include <mutex>
 
-#include <c10/util/Exception.h>
 #include <c10/util/SmallVector.h>
 #include <c10/util/flat_hash_map.h>
 

c10/util/flags_use_gflags.cpp

@@ -1,11 +1,10 @@
-#include <c10/util/Flags.h>
-
-#include <string>
-
 #include <c10/macros/Macros.h>
 
 #ifdef C10_USE_GFLAGS
 
+#include <c10/util/Flags.h>
+#include <string>
+
 namespace c10 {
 
 using std::string;

c10/util/signal_handler.cpp

@@ -1,18 +1,16 @@
 #include <c10/util/Backtrace.h>
+#include <c10/util/Logging.h>
 #include <c10/util/signal_handler.h>
 
 #if defined(C10_SUPPORTS_SIGNAL_HANDLER)
 
 // Normal signal handler implementation.
-#include <cxxabi.h>
 #include <dirent.h>
-#include <dlfcn.h>
-#include <fmt/format.h>
+#include <fmt/core.h>
 #include <sys/syscall.h>
-#include <sys/types.h>
 #include <unistd.h>
-#include <unwind.h>
 
+#include <cstdint>
 #include <cstdio>
 #include <cstdlib>
 #include <iostream>

c10/util/signal_handler.h

@@ -5,7 +5,6 @@
 #include <mutex>
 
 #include <c10/macros/Export.h>
-#include <c10/util/Logging.h>
 
 #if defined(__APPLE__)
 #define C10_SUPPORTS_SIGNAL_HANDLER