Optimize tm formatting (Non C-locales and %Z) (#2617)

* Move fmt::detail::formatbuf to format.h

* Replace std::basic_ostringstream to std::basic_ostream with custom formatbuf

* Use tm.tm_zone

include/fmt/format.h

@@ -265,6 +265,38 @@ FMT_END_NAMESPACE
 
 FMT_BEGIN_NAMESPACE
 namespace detail {
+
+template <typename Streambuf> class formatbuf : public Streambuf {
+ private:
+  using char_type = typename Streambuf::char_type;
+  using streamsize = decltype(std::declval<Streambuf>().sputn(nullptr, 0));
+  using int_type = typename Streambuf::int_type;
+  using traits_type = typename Streambuf::traits_type;
+
+  buffer<char_type>& buffer_;
+
+ public:
+  explicit formatbuf(buffer<char_type>& buf) : buffer_(buf) {}
+
+ protected:
+  // The put area is always empty. This makes the implementation simpler and has
+  // the advantage that the streambuf and the buffer are always in sync and
+  // sputc never writes into uninitialized memory. A disadvantage is that each
+  // call to sputc always results in a (virtual) call to overflow. There is no
+  // disadvantage here for sputn since this always results in a call to xsputn.
+
+  auto overflow(int_type ch) -> int_type override {
+    if (!traits_type::eq_int_type(ch, traits_type::eof()))
+      buffer_.push_back(static_cast<char_type>(ch));
+    return ch;
+  }
+
+  auto xsputn(const char_type* s, streamsize count) -> streamsize override {
+    buffer_.append(s, s + count);
+    return count;
+  }
+};
+
 // An equivalent of `*reinterpret_cast<Dest*>(&source)` that doesn't have
 // undefined behavior (e.g. due to type aliasing).
 // Example: uint64_t d = bit_cast<uint64_t>(2.718);