fix: gumbo memory leak on abandoned tags (#3036)

**What problem is this PR intended to solve?**

The fuzzer in #3007 found a memory leak when incomplete tags are
encountered.

The repro added to the memory leak test suite is:

```html
<asdfasdfasdfasdfasdfasdfasdfasdfasdfasdf foo="bar
```

which will leak the tag name without this fix.

cc @stevecheckoway 

**Have you included adequate test coverage?**

Yes, but added it to the memory leak suite which can only be run
manually today (see #1603 for context on plans to improve it).

**Does this change affect the behavior of either the C or the Java
implementations?**

No functional change.

gumbo-parser/src/tokenizer.c

@@ -506,6 +506,7 @@ static void abandon_current_tag(GumboParser* parser) {
   for (unsigned int i = 0; i < tag_state->_attributes.length; ++i) {
     gumbo_destroy_attribute(tag_state->_attributes.data[i]);
   }
+  gumbo_free(tag_state->_name);
   gumbo_free(tag_state->_attributes.data);
   mark_tag_state_as_empty(tag_state);
   gumbo_string_buffer_destroy(&tag_state->_buffer);

test/test_memory_leak.rb

@@ -270,6 +270,20 @@ def test_leaking_dtd_nodes_after_internal_subset_removal
         puts
       end
     end
+
+    describe "libgumbo abandoned tag" do
+      it "should not leak the tag name" do
+        html = <<~HTML
+          <asdfasdfasdfasdfasdfasdfasdfasdfasdfasdf foo="bar
+        HTML
+        # should increase over the first 200_000 iterations (general parsing overhead),
+        # but then flatten out. on my machine at about 169k
+        1_000_000.times do |j|
+          Nokogiri::HTML5::Document.parse(html)
+          printf "%s::%s: (iter %d) %d\n", self.class, __method__, j, MemInfo.rss if j % 20_000 == 0
+        end
+      end
+    end
   end # if NOKOGIRI_GC
 
   def test_object_space_memsize_of
@@ -336,7 +350,8 @@ module MemInfo
     rescue
       4096
     end
-    STATM_PATH = "/proc/#{Process.pid}/statm"
+
+    STATM_PATH = "/proc/self/statm"
     STATM_FOUND = File.exist?(STATM_PATH)
 
     def self.rss